Estimation of Regional Electricity Consumption Using National Polar-Orbiting Partnership’s Visible Infrared Imaging Radiometer Suite Night-Time Light Data with Gradient Boosting Regression Trees

Abstract

With the rapid development of society and economy, the growth of electricity consumption has become one of the important indicators to measure the level of regional economic development. This paper utilizes NPP-VIIRS nighttime light remote sensing data to model electricity consumption in parts of southern China. Four predictive models were initially selected for evaluation: LR, SVR, MLP, and GBRT. The accuracy of each model was assessed by comparing real power consumption with simulated values. Based on this evaluation, the GBRT model was identified as the most effective and was selected to establish a comprehensive model of electricity consumption. Using the GBRT model, this paper analyzes electricity consumption in the study area across different spatial scales from 2013 to 2022, demonstrating the distribution characteristics of electricity consumption from the pixel level to the city scale and revealing the close relationship between electricity consumption and regional economic development. Additionally, this paper examines trends in electricity consumption across various temporal scales, providing a scientific basis for the optimal allocation of energy and the effective distribution of power resources in the study area. This analysis is of great significance for promoting balanced economic development between regions and enhancing energy efficiency.

1. Introduction

Electricity consumption is a crucial indicator of socioeconomic development, with rapid economic growth typically accompanied by high electricity demand. As China’s economy has developed, the patterns of electricity consumption have also transformed, characterized by adjustments in consumption structure and significant regional disparities [1]. Consequently, analyzing electricity consumption data becomes particularly important when assessing the power demand and economic status of a region [2]. However, there are two main challenges in obtaining electricity consumption data: firstly, the granularity of the data is insufficient, failing to reach the county, town, or more micro levels [3]; secondly, the update speed of electricity statistics is slow, with some regions experiencing data gaps [4]. Therefore, it is imperative to develop a method that can quickly and accurately reflect regional electricity consumption. Such a method would effectively improve the allocation of electrical resources, promote balanced regional economic development, and enhance energy efficiency.

Remote sensing can provide large spatial extent and long-term imagery, forming the data foundation for simulating regional power consumption [5]. Satellite sensors capture nighttime visible light radiation sources, such as urban lights, fishing boat illumination, and fires under cloud-free conditions, generating nighttime light (NTL) remote sensing images [6,7]. Nighttime light remote sensing has been widely applied in extracting urban built-up areas [8,9], spatializing population distribution [10,11], and spatializing economic parameters [12,13,14], with particularly outstanding performance in power consumption inversion [15,16,17,18]. The new generation of nighttime light data product, NPP-VIIRS (National Polar-orbiting Partnership’s Visible Infrared Imaging Radiometer Suite), effectively overcomes the limitations of DMSP-OLS (Defense Meteorological Satellite Program Operational Linescan System) data in terms of spatial, temporal, and radiometric resolution [19]. It enables fine observations of smaller-scale areas and effectively avoids pixel brightness saturation, thus enhancing observational sensitivity. Existing research results show that, compared to DMSP-OLS nighttime light data, NPP-VIIRS demonstrates better accuracy in estimating power consumption [20,21].

In the past decade, numerous studies have shown that nighttime light remote sensing data have a strong linear relationship with power consumption. Shi et al. calibrated global power consumption data from DMSP-OLS nighttime light remote sensing data from 1992 to 2013 and conducted an analysis of the spatiotemporal dynamics of power consumption at various global scales [22]. Chen et al. used a particle swarm optimization-back propagation algorithm to unify DMSP/OLS and NPP/VIIRS data, generating nighttime remote sensing products at a 1 km resolution from 1992 to 2019, and derived a global-scale power consumption dataset [23]. Sahoo et al. estimated power consumption in northern India using NPP-VIIRS and DMSP-OLS nighttime light data and developed a power consumption estimation product based on NPP-VIIRS nighttime light remote sensing data from 2013 to 2017. The research results indicated that using principal component analysis could achieve more accurate estimation results, and the estimates generated from NPP-VIIRS data were significantly better than those from DMSP-OLS [20]. Zhong used power consumption data inverted from DMSP-OLS nighttime light data and analyzed the spatiotemporal dynamic evolution characteristics of power consumption in three national-level urban agglomerations in the Yangtze River Economic Belt of China (Yangtze River Delta urban agglomeration, middle reaches urban agglomeration, and Chengdu-Chongqing urban agglomeration) using various spatial analysis methods. They also analyzed the factors influencing power consumption using the geographically weighted regression model and random forest algorithm [24]. More recent studies have leveraged newer nighttime light datasets, such as SDGSAT-1, which provides higher spatial resolution. SDGSAT-1 imagery has been applied in evaluating road network power conservation [25] and investigating the spatial variability of nighttime lights across urban areas [26], contributing to further insights in this field. These studies complement the broader body of research by offering additional datasets for analyzing power consumption.

Additionally, some researchers proposed a novel classification regression method using nighttime light remote sensing images to improve the accuracy of power consumption estimation in China. This method, which incorporates spatially nonstationary relationships between nighttime light data and electric power consumption, was shown to outperform traditional models in terms of relative error and mean absolute percentage error [27]. Another study developed a model using saturation-corrected DMSP-OLS nighttime light data to estimate the spatiotemporal dynamics of electric power consumption in China from 2000 to 2008. This model demonstrated high reliability with an average R value of 0.93, showing that it could effectively capture regional variations in power consumption [28]. Ren et al. utilized a parcel-oriented temporal linear unmixing method to identify specific nighttime light sources and improve the estimation accuracy of power consumption in Shanghai, revealing the strong correlation between sectoral energy consumption and unmixed nighttime light data [29]. Jin et al. applied a random forests model integrating points of interest (POIs) and multiple remote sensing data to produce high-resolution maps of industrial and nonindustrial electric power consumption across China, achieving the high estimation accuracy [30].

While methods for estimating electricity consumption using nighttime light remote sensing data have achieved considerable accuracy, comprehensive analyses that explore electricity consumption changes across multiple spatiotemporal scales are still limited. In this study, we utilize NPP-VIIRS nighttime remote sensing data to construct a detailed model of electricity consumption at the city level. To ensure the most accurate representation, we compared several predictive models and identified the most effective one for our analysis. This optimal model was then applied to extend our study to pixel (500 m × 500 m) and city levels, allowing for a comprehensive exploration of electricity consumption patterns across different spatiotemporal scales in the parts of southern China. This region is of particular importance due to its rapid economic development, high population density, and significant industrial activities, making it a critical area for energy consumption analysis. Understanding and analyzing the electricity consumption patterns in the study area is essential for optimizing energy allocation, promoting sustainable development, and formulating effective energy policies [31,32].

2. Materials and Methods

2.1. Study Area

This study focuses on parts of southern China, encompassing Guangdong Province, Fujian Province, and Guangxi Zhuang Autonomous Region (Figure 1). The total land area of South China is approximately 538,800 km², featuring diverse terrain including mountains, hills, and plains. The region boasts a coastline extending about 9124 km along the South China Sea, with a sea area of 3.5 million km². The study area’s climate is characterized by tropical and subtropical monsoon patterns. Guangdong Province, as the largest economic powerhouse in the region, recorded a gross domestic product (GDP) of RMB 13.57 trillion (approximately USD 1.89 trillion) in 2023, accounting for about one-tenth of the national total. The province’s economy grew by 4.8% year-on-year, driven by its robust manufacturing sector, which represents 32.7% of its GDP. Guangdong is home to over 71,000 industrial enterprises with annual revenues exceeding RMB 20 million and more than 75,000 high-tech enterprises.

Fujian Province and Guangxi Zhuang Autonomous Region also significantly contribute to the regional economy. In 2023, Fujian’s GDP reached RMB 5.3 trillion (approximately USD 744 billion), reflecting steady economic growth. Similarly, Guangxi’s GDP was RMB 2.72 trillion (approximately USD 380 billion), with a year-on-year growth rate of 4.1%. The demographic composition of the study area underscores its importance, with Guangdong, the most populous province in China, having approximately 126 million people. Fujian and Guangxi have populations of about 39 million and 49 million, respectively. Analyzing power consumption patterns and trends in the study area is critical for understanding the region’s economic development and for informing national energy strategies. The unique economic, geographic, and climatic characteristics of the study area provide a compelling case for a multispatiotemporal analysis of power consumption, aiming to offer essential insights for optimizing electricity usage and supporting the sustainable and healthy development of the region’s economy.

2.2. Data Sources

The nighttime light remote sensing is sourced from the monthly light product synthesized by the Colorado School of Mines based on NPP-VIIRS data, available at “https://eogdata.mines.edu/products/vnl/ (accessed on 12 October 2024)”, covering the period from 2013 to 2022 [33,34]. The monthly synthesized data filter out data from auroras, fires, ships, and other temporary light sources. The nighttime remote sensing data used in this paper come from the VIIRS sensor on the Suomi NPP satellite, which carries five sensors, including VIIRS, and was successfully launched on 28 October 2011. The Suomi NPP satellite crosses its orbit 14 times daily, with local crossing times at the equator of 1:30 am and 1:30 pm. The VIIRS sensor has 22 spectral channels, with the nighttime remote sensing utilizing the day/night band (DNB), which has a spectral range of 500–900 nm and a spatial resolution of 750 m.

The city electricity consumption data (total electricity consumption) used in this study are sourced from the “China City Statistical Yearbook”, which is published and compiled by the Urban Social and Economic Investigation Department of the National Bureau of Statistics. Total electricity consumption includes the total energy consumption across all sectors that use electricity within a specific region, encompassing the primary, secondary, and tertiary industries. This comprehensive metric objectively reflects the electricity usage patterns of the area. The data are collected at the city level, span the years 2018 to 2020, and have a temporal resolution of one year. In this paper, we consider the statistical data released by the National Bureau of Statistics to be the authentic and reliable data. Based on these data, we conducted modeling and validation of electricity consumption. In the context of this study, “city” refers to administrative units in China, which often include both urban and rural areas within the same jurisdiction.

2.3. Methods

The NPP-VIIRS nighttime light data from 2013 to 2022 are utilized to estimate the electricity consumption at both city and grid levels in the study area, following the three main procedures outlined in our methodology (Figure 2). The first part involves data preprocessing, where the NPP-VIIRS monthly nighttime light remote sensing data are converted into an annual scale dataset. The second part focuses on model development. Four machine learning models, namely, linear regression (LR), support vector regression (SVR), multilayer perceptron (MLP), and gradient boosting regression trees (GBRT), are employed to establish the relationship between the nighttime light data and city-level electricity consumption. These models are trained using the annual nighttime light data as the independent variable and the corresponding yearly electricity consumption data as the dependent variable. To identify the most robust and accurate model, a 5-fold cross-validation process is applied. This method involves dividing the data into five subsets, training the model on four subsets, and validating it on the remaining one in an iterative manner. The cross-validation ensures that the selected model generalizes well to unseen data, minimizing the risk of overfitting.

In the final part, the optimal model, determined from the cross-validation process, is used to calculate electricity consumption at both the city and grid levels. For city-level estimates, the model directly applies the relationship established during the training phase. However, estimating electricity consumption at the pixel level (grid scale) requires a more nuanced approach, which is achieved through the weight allocation method. The weight allocation method plays a critical role in refining the pixel-level electricity consumption estimates. This technique involves assigning weights to the contributions of different pixels based on their relative significance in the context of the overall electricity consumption model. The core principle of weight allocation is to distribute the total predicted electricity consumption across individual pixels in a way that reflects their respective intensities of nighttime light. Mathematically, this can be expressed as

$$P_i={\displaystyle \frac{W_i\times P_{total}}{{\displaystyle {\sum }_{i=1}^n{W}_i}}},$$

(1)

where $P_i$ represents the estimated electricity consumption for pixel i, $W_i$ is the electricity consumption predicted by the optimal model for pixel i, and $P_{total}$ is the total electricity consumption predicted by the optimal model for the entire area. This weight allocation method ensures that the nonlinear relationships modeled at the city level can be effectively applied and extended to the grid level, enabling more accurate estimations of electricity consumption at finer spatial resolutions. By integrating this method into the overall estimation process, this study achieves a detailed and accurate depiction of electricity consumption patterns at both the city and grid levels.

2.3.1. Data Preprocessing

Due to the high radiometric resolution of the VIIRS sensor, it can detect both very high and very low levels of light, resulting in background noise and outliers in the images. Although the NPP-VIIRS monthly composite data were corrected for factors such as cloud cover, they can still be affected by noise from unstable light sources or stray light. To improve the quality of the pixels, it is necessary to remove this noise and enhance the continuity between images through interannual correction.

First, the light values in areas covered by large bodies of water are set to zero. Then, an empirical threshold method is used to process sporadic weak lights (background noise) in the NPP-VIIRS data by setting digital number (DN) values below 0.5 to zero. For extremely high noise values, a similar threshold 0 is applied, where pixels exceeding this threshold are eliminated. The method for selecting the high threshold is as follows: the maximum value of nighttime light remote sensing pixels should be distributed in the most economically developed commercial core areas. If pixel values in other regions exceed this maximum value, they are considered outliers and are set to no data. In this study, Shenzhen, the most economically developed city in the study area, is used as the reference city. The maximum pixel values from all years in Shenzhen are used as the high threshold.

After correcting the nighttime light remote sensing data, to reduce map distortion, the data are reprojected from the WGS84 geographic coordinate system to the Lambert equal-area projection coordinate system and resampled to 500 m. The monthly data are summed to obtain the annual data. The nighttime light remote sensing data are then clipped according to provincial and municipal administrative boundaries to produce corresponding datasets for each province and city.

2.3.2. Simulation of Energy Consumption

The total nighttime light (TNL) for each city within the study area is calculated based on nighttime remote sensing data. The calculation formula is as follows:

$$TNL=\sum _{i=1}^n{x}_i,$$

(2)

TNL represents the total nighttime light for the city, and $x_i$ is the brightness value of the i pixel within the city’s area. After obtaining the total nighttime light for each city, we apply four different modeling techniques to establish the relationship between total electricity consumption and total nighttime light. These methods include LR, SVR, MLP, and GBRT. Each model is assessed, and the best-performing model is selected based on its accuracy and generalizability.

Linear regression (LR) is a fundamental statistical method used to model the relationship between a dependent variable and one or more independent variables. In this study, LR is utilized to establish the relationship between TNL data derived from NPP-VIIRS imagery and city-level electricity consumption. The model assumes a linear relationship, meaning that as the TNL increases, electricity consumption increases proportionally. The general formula for the LR model is

$$y={\beta }_0+{\beta }_1\times TNL+ϵ,$$

(3)

where y represents the total electricity consumption, ${\beta }_0$ is the intercept, ${\beta }_1$ is the slope indicating the rate of change in electricity consumption per unit change in TNL, and $ϵ$ is the error term accounting for any variation not explained by the model. LR is widely used due to its simplicity and ease of interpretation, making it an ideal starting point for understanding the relationship between TNL and electricity consumption. However, it assumes a linear relationship, which may not fully capture the complexities of the data, particularly in cases where the relationship is nonlinear [35].

Support vector regression (SVR) is an extension of support vector machines (SVMs) used for regression tasks. SVR is particularly effective in modeling nonlinear relationships by mapping the input data into a higher-dimensional space using a kernel function. The SVR model aims to find a function that deviates from the observed electricity consumption by no more than a specified margin $ϵ$, while also ensuring that the model is as flat as possible. The general form of the SVR model is

$$y=w\times TNL+b,$$

(4)

subject to $|y-(w\times TNL+b)|\le ϵ$, where w and b are coefficients determined by the model, and $ϵ$ is the margin of tolerance. SVR is particularly robust in handling nonlinear patterns and is less sensitive to outliers, making it a suitable choice for complex datasets where linear models may not perform well. However, SVR requires careful tuning of hyperparameters and the choice of an appropriate kernel to achieve optimal performance [36].

The multilayer perceptron (MLP) is a type of artificial neural network that consists of an input layer, one or more hidden layers, and an output layer. Each layer contains nodes (neurons) that process the input data through a series of weighted connections and activation functions. MLP is highly capable of modeling complex, nonlinear relationships between TNL and electricity consumption. The general formula for the MLP model is

$$y=\sigma \left(\sum _{j=1}^m{w}_j\times \sigma \left(\sum _{i=1}^n{w}_{ij}\times TN{L}_i+b_i\right)+b_j\right),$$

(5)

where $w_{ij}$ are the weights of the connections, $b_i$ and $b_i$ are the biases, and $\sigma$ is the activation function (e.g., ReLU, sigmoid) that introduces nonlinearity into the model. MLP is particularly powerful in capturing complex interactions within the data, but it requires substantial computational resources and is prone to overfitting if not properly regularized [37].

Gradient boosting regression trees (GBRT) is an ensemble learning method that builds a model in a stage-wise manner by combining the predictions of multiple weak learners, typically decision trees. Each new tree is trained to correct the errors made by the previous trees, gradually improving the model’s accuracy. The formula for the GBRT model is

$$y=\sum _{k=1}^K{\alpha }_k{T}_k(TNL),$$

(6)

where $T_k(TNL)$ represents the prediction from the k-th tree, ${\alpha }_k$ is the learning rate or weight of the k-th tree, and K is the total number of trees in the ensemble. GBRT is highly effective in capturing both linear and nonlinear relationships and is known for its robustness and high accuracy, especially in cases where the data contain complex patterns and interactions [38].

2.3.3. Accuracy Evaluating

Model validation is a vital step in evaluating the performance of a model. To determine the effectiveness of the least squares method’s estimations, this study employed three metrics: the coefficient of determination (R²), the relative deviation (RD), and the root mean square error (RMSE), corresponding to Equations (7)–(9). R² measures how well the model’s estimates explain the actual values, also referred to as goodness of fit. It ranges from 0 to 1, with values closer to 1 indicating better model accuracy. RD reflects the proportion of deviation between estimated and actual values, representing the relative error in percentage form. A lower RD value signifies higher model accuracy. RMSE indicates the average difference between estimated and actual values. A smaller RMSE value suggests better model performance by indicating a smaller average deviation.

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{j=1}^n}{\left(E_j-{\widehat{E}}_j\right)}^2}{{\displaystyle \sum _{j=1}^n}{\left(E_j-\overline{E}\right)}^2}},$$

(7)

$$RD={\displaystyle \frac{1}{n}}\sum _{j=1}^n\left|{\displaystyle \frac{E_j-{\widehat{E}}_j}{E_j}}\right|\times 100\%,$$

(8)

$$RMSE=\sqrt{\frac{1}{n}\sum _{j=1}^n{\left(E_j-{\hat{E}}_j\right)}^2},$$

(9)

where $E_j$ is the actual total electricity consumption for the j-th city, ${\widehat{E}}_j$ is the estimated total electricity consumption for the j-th city, $\overline{E}$ is the mean of actual total electricity consumption values, and n is the number of cities.

3. Results

3.1. Simulation Results of Different Models

The performance of the four models (LR, SVR, MLP, and GBRT) was evaluated by generating scatter plots that compare the simulated total electricity consumption with the actual total electricity consumption (Figure 3). The LR model yielded an R² value of 0.867, with an RD of 26.64% and an RMSE of 56.32 GWh. This model demonstrates a reasonable correlation between the simulated and actual electricity consumption, though the scatter of points around the regression line, particularly at higher consumption levels, suggests that LR may not fully capture the complexity of the relationship between nighttime light data and electricity consumption. The SVR model, illustrated in the second subplot, shows improved performance with an R² value of 0.886, an RD of 21.8%, and an RMSE of 42.2 GWh. The tighter clustering of points around the regression line indicates that SVR handles nonlinear relationships more effectively, particularly in the mid-range consumption levels. The MLP model achieved an R² value of 0.880, an RD of 16.6%, and an RMSE of 45.5 GWh. While MLP improves upon the correlation between simulated and actual values, some dispersion in the predictions indicates that while it captures more complex patterns, it may also introduce variability that impacts prediction accuracy. The GBRT model, represented in the fourth subplot, outperforms all others, with an R² value of 0.905, an RD of 13.9%, and an RMSE of 30.6 GWh, demonstrating the highest level of accuracy among the models, with closely clustered points around the regression line. The regression lines also show a tendency to overestimate consumption in areas with high electricity demand and underestimate it in areas with lower demand. This is likely due to the data-driven nature of the models, where cities with extremely high or low electricity consumption are underrepresented in the dataset. As a result, the models perform well on mid-range consumption values but may struggle with generalization at the extremes, causing the observed deviations.

While the OLS model serves as a useful baseline for comparison, it may not fully capture the nonlinear relationships between nighttime light and electricity consumption. The machine learning approaches, especially GBRT, have the flexibility to model these more complex interactions. The superior performance of the GBRT model can be attributed to its ability to incrementally correct errors from previous iterations, refining the prediction model through its ensemble approach. This allows GBRT to effectively capture complex patterns and interactions within the data, leading to more precise and reliable estimates of electricity consumption. Its success in outperforming the other models highlights the advantages of using ensemble methods in predictive modeling, particularly for datasets that exhibit both linear and nonlinear characteristics.

Furthermore, the analysis reveals significant differences in electricity consumption across the provinces. Guangdong Province, being more economically developed, shows higher electricity consumption in most cities. Fujian Province has a moderate level of electricity consumption, whereas Guangxi Zhuang Autonomous Region, with its relatively weaker economy, displays noticeably lower electricity consumption compared to the other two provinces. These findings underscore the substantial spatial variations in electricity consumption across the region. Given the superior performance of the GBRT model, we utilized it to simulate electricity consumption at various scales, providing insights into the consumption patterns across different spatial dimensions. This model’s ability to accurately reflect the regional disparities in consumption further validates its application in analyzing electricity usage at both city and grid levels.

3.2. Spatial Analysis of Electricity Consumption

3.2.1. Grid Scale

Using the preprocessed nighttime light remote sensing data, along with the GBRT model and weight allocation method, we can estimate the electricity consumption at the grid scale for the study area for the years 2013 to 2022. The results are shown in Figure 4, where four selected years are presented for illustration. As illustrated, Guangdong Province exhibits higher electricity consumption, particularly in the Pearl River Delta region, which is the main concentration area for electricity usage. In Fujian Province, electricity consumption is primarily concentrated in coastal areas such as Xiamen, Fuzhou, and Wenzhou, while inland areas show lower electricity consumption. In Guangxi Zhuang Autonomous Region, electricity consumption is mainly focused in inland areas such as Nanning and Liuzhou, with the average annual electricity consumption over the past decade being 25.36 GWh and 18.48 GWh, respectively.

The spatial distribution of electricity consumption shows distinct patterns. In areas of low electricity consumption, such as Guangxi Zhuang Autonomous Region and Fujian Province, the consumption is mainly spotty, concentrated in economically developed urban areas. In contrast, areas of high electricity consumption, such as the Pearl River Delta, exhibit a network-like distribution, where multiple cities are interconnected, forming an agglomeration effect.

These spatial patterns indicate that economic development significantly influences electricity consumption. In high consumption areas, the interconnected urban networks facilitate higher energy usage, whereas in lower consumption areas, the usage is more localized and concentrated in specific urban centers. This analysis underscores the importance of considering spatial heterogeneity when planning for energy resource allocation and infrastructure development in the study area region.

3.2.2. City Scale

Using a classification method, the city-level electricity consumption in the study area region from 2013 to 2022 was divided into five categories: 50–200 GWh, 200–400 GWh, 400–600 GWh, 600–800 GWh, and >800 GWh. The spatial distribution of electricity consumption at the city scale in the study area is shown in Figure 5. As depicted in Figure 4, the overall electricity consumption in the region shows an increasing trend.

In 2016, electricity consumption peaked, with four cities—Guangzhou, Dongguan, Shenzhen, and Quanzhou—having annual electricity consumption exceeding 800 GWh. In contrast, in 2013, only Guangzhou reached the 600–800 GWh level, while Shenzhen, Dongguan, and Quanzhou were in the 400–600 GWh range. By 2019, electricity consumption in the study area had slightly declined, with Guangzhou being the highest at over 800 GWh, followed by Shenzhen in the 600–800 GWh range. By 2022, the number of cities with electricity consumption levels of 600–800 GWh had returned to three, namely, Shenzhen, Dongguan, and Quanzhou.

During the period from 2013 to 2022, there was a notable decrease in the number of cities with electricity consumption levels of 50–200 GWh, dropping from 26 cities in 2013 to 18 cities in 2022. Compared to the peak in 2016, the number of cities with electricity consumption levels of 50–200 GWh decreased by four in 2022. This indicates that the economic development among cities within the study area is gradually becoming more balanced. The analysis of city-scale electricity consumption from 2013 to 2022 reveals significant spatial and temporal variations. The observed trends highlight the dynamic nature of economic growth and energy usage, underscoring the importance of continuous monitoring and tailored energy policies to address the evolving needs of different urban areas.

3.3. Temporal Analysis of Electricity Consumption

By compiling and analyzing the annual electricity consumption data of provinces and autonomous regions, we constructed a line chart depicting the annual electricity consumption at the provincial level from 2013 to 2022 (Figure 6). The average annual electricity consumption over this period for Guangdong Province, Fujian Province, and Guangxi Zhuang Autonomous Region was 479.082 GWh, 208.980 GWh, and 158.482 GWh, respectively, with Guangdong Province significantly outpacing the other two provinces.

Figure 6 clearly shows that Guangdong Province maintained a consistently higher electricity consumption than the other provinces over the ten-year period, with Fujian Province slightly exceeding Guangxi Zhuang Autonomous Region. From a temporal perspective, the annual electricity consumption in the study area exhibited an overall increasing trend over the past decade, peaking between 2015 and 2016. In 2016, Guangdong and Fujian provinces reached their highest consumption levels at 778.719 GWh and 334.829 GWh, respectively, while Guangxi Zhuang Autonomous Region peaked in 2020 at 250.165 GWh. A notable drop in electricity consumption was observed in 2017 across the study area, followed by a gradual recovery in subsequent years. This analysis underscores the dynamic nature of electricity consumption, influenced by economic growth, industrial activity, and regional development policies.

To further analyze electricity consumption trends, we examined the multiyear average monthly electricity consumption data for each province and autonomous region, and constructed a line chart illustrating the monthly variations at the provincial level in the study area (Figure 7). The trends of monthly electricity consumption across different provinces and autonomous regions in the study area are largely consistent, with relatively small variations in magnitude. In Guangdong and Fujian provinces, the lowest electricity consumption occurs in July, with values of 35.804 billion kWh and 16.003 billion kWh, respectively. In contrast, Guangxi Zhuang Autonomous Region experiences its lowest electricity consumption in January, with a value of 10.377 billion kWh. The highest monthly electricity consumption for all three provinces/regions occurs in October, with Guangdong, Fujian, and Guangxi Zhuang Autonomous Region reaching 44.002 billion kWh, 19.490 billion kWh, and 16.635 billion kWh, respectively.

These findings highlight the seasonal patterns in electricity consumption, which are influenced by factors such as climate, industrial activity, and population behavior. The observed consistency in monthly consumption trends suggests a region-wide response to these factors, although the specific timing and magnitude of peaks and troughs can vary due to local conditions. However, it is important to note that nighttime light remote sensing primarily captures persistent light sources, such as street lighting and certain industrial operations, and may not fully reflect electricity consumption driven by daytime or seasonal variations like heating, cooling, or fluctuating industrial activities. Understanding these patterns is essential for optimizing energy resource management and ensuring reliable electricity supply throughout the year.

3.4. Analysis of Electricity Consumption in Typical Cities of the Study Area

Based on the simulation of electricity consumption in the study area, this study selects four representative cities—Guangzhou, Shenzhen, Fuzhou, and Nanning—to analyze the spatiotemporal changes in their electricity consumption. Among these, Nanning, Fuzhou, and Guangzhou are the capital cities of their respective provinces or autonomous regions.

Figure 8 illustrates the spatial distribution characteristics of electricity consumption in Nanning and Fuzhou for the years 2013 and 2022. As shown in the figure, electricity consumption in Nanning is primarily concentrated in the southwestern part of the city, while other areas exhibit a scattered pattern of consumption. Over time, by 2022, the core consumption area in Nanning has significantly expanded, with new pockets of electricity consumption emerging particularly in the northwestern region. In contrast, Fuzhou’s main electricity consumption areas are located in the eastern coastal region. By 2022, the concentrated consumption zones in Fuzhou have notably expanded, and there is an increasing trend in electricity consumption on offshore islands. These spatial expansions in electricity consumption areas reflect the economic development and urbanization processes occurring in these cities.

As the top two economic powerhouses in Guangdong Province, Guangzhou and Shenzhen are both critical cities in the Guangdong–Hong Kong–Macau Greater Bay Area. Figure 9 reveals the spatial distribution characteristics of electricity consumption in these two cities for the years 2013 and 2022. Shenzhen exhibits distinct electricity consumption characteristics, with most areas displaying high consumption levels, except for relatively lower consumption in the southeastern region. By 2022, electricity consumption had increased across all areas of Shenzhen, particularly in the northwestern region, where the deepened color indicates significant growth in electricity usage. The generally high electricity consumption levels throughout Shenzhen are primarily due to its relatively small urban area (1996.78 square kilometers).

In contrast, Guangzhou has a much larger urban area of 7434.4 square kilometers. The core electricity consumption area in Guangzhou is mainly located in the southwestern part of the city. From 2013 to 2022, the spatial distribution pattern of electricity consumption in Guangzhou transitioned from a network-like pattern to a more contiguous area, indicating an overall increase in electricity consumption. This change reflects the substantial development Guangzhou has achieved over the past decade.

Figure 10 presents the electricity consumption trends for Guangzhou, Shenzhen, Fuzhou, and Nanning from 2013 to 2022. It reveals the annual changes in electricity consumption for each city. Guangzhou leads the other cities with an average annual consumption of 91.413 GWh, followed by Shenzhen (64.688 GWh), Fuzhou (49.422 GWh), and Nanning (39.866 GWh). Over the ten-year period, all four cities exhibit significant synchronization in their consumption trends.

The electricity consumption in these major cities of the study area reached its lowest point in 2014, with Guangzhou consuming 60.535 GWh, Shenzhen 44.775 GWh, Fuzhou 31.251 GWh, and Nanning 22.691 GWh. Subsequently, Guangzhou, Shenzhen, and Fuzhou experienced their peak consumption in 2016, reaching 128.046 GWh, 91.225 GWh, and 67.935 GWh, respectively, indicating significant growth. In contrast, Nanning’s peak electricity consumption occurred in 2020, with a total of 48.753 GWh.

4. Discussion

4.1. Comparison with Previous Studies

Nighttime light (NTL) data have been extensively used to estimate electricity consumption and other socioeconomic indicators. Previous studies have demonstrated a strong correlation between NTL data and electricity consumption, highlighting the potential of this method for various applications. For example, Shi et al. compared the performance of DMSP-OLS and NPP-VIIRS data in modeling electricity consumption and found that NPP-VIIRS provided higher accuracy due to its better spatial resolution and wider radiometric detection range [13]. Similarly, Zhu et al. used causal-effect inference to test the suitability of NTL data for estimating electricity consumption and concluded that it was more appropriate for developing countries, while developed countries required the inclusion of more latent factors [38]. Chen et al. addressed the limitations of overestimating real GDP growth and the heterogeneity in spatiotemporal dynamics by developing a global 1 km × 1 km gridded dataset of electricity consumption for 1992–2019 using a particle swarm optimization-back propagation (PSO-BP) algorithm to unify DMSP-OLS and NPP-VIIRS data [23]. Hu et al. further enhanced this approach by integrating DMSP-OLS and NPP-VIIRS nighttime light data to produce a consistent pixel-level electricity consumption product spanning 1992 to 2019, providing accurate estimates across global, continental, and national scales [39].

Comparatively, this study’s approach aligns with previous findings by utilizing NPP-VIIRS data to achieve high accuracy in estimating electricity consumption across different spatial scales. However, our study extends the analysis to a more granular level, including pixel and provincial scales, and incorporates a comprehensive temporal range from 2013 to 2022. This extension allows for a more detailed understanding of regional consumption patterns and their temporal evolution.

Moreover, recent advancements have explored integrating other data sources with NTL data to improve estimation accuracy. Sun et al. combined demographic, remote sensing, and social sensing data to estimate electricity consumption at a local scale, demonstrating the enhanced explanatory power of such mixed approaches [40]. Similarly, Deng et al. used smart meter readings to track electricity consumption at the pixel level, addressing the coarse resolution limitations of traditional NTL imagery [41]. These studies highlight the potential for integrating multiple data sources to achieve higher resolution and accuracy in electricity consumption estimation, suggesting future research directions for improving our models.

4.2. Implications for Energy Planning and Policy

Understanding electricity consumption patterns through nighttime light data provides significant implications for energy planning and policy development. The ability to accurately estimate electricity consumption enables more informed decisions regarding the allocation of energy resources, which can, in turn, enhance energy efficiency and support sustainable economic growth. For example, NTL data allow for real-time monitoring of energy use across different regions, offering a granular understanding of consumption patterns that can guide the strategic placement of energy infrastructure. This approach is particularly valuable in rapidly developing urban areas, where energy demands are constantly shifting. Previous studies, such as the work by Guo et al., have demonstrated the utility of NPP-VIIRS data in analyzing the spatiotemporal dynamics of electricity consumption in Xi’an, China, providing valuable insights for urban planning and the rational allocation of electric power resources [42].

The findings of this study, which highlight significant spatial variations in electricity consumption across different provinces in South China, can serve as a critical tool for policymakers. By identifying regions with disproportionately high or low energy demands, targeted interventions can be designed to optimize energy distribution and reduce regional disparities. For instance, the high electricity consumption observed in Guangdong Province—an economically advanced region with extensive industrial activities—suggests the need for energy policies focused on enhancing efficiency within the industrial sector. This could include the implementation of energy-saving technologies, stricter regulations on industrial emissions, and incentives for the adoption of renewable energy sources. Conversely, the lower consumption levels noted in Guangxi Zhuang Autonomous Region, which has a relatively weaker economy, point to the potential for further economic development through investment in energy infrastructure. Strategic investments in these regions could not only bolster local economies but also ensure that energy supply meets future demands as these areas continue to develop.

Additionally, the temporal analysis of electricity consumption, which reveals clear seasonal variations, underscores the importance of dynamic and responsive energy planning. Seasonal peaks and troughs in electricity demand, driven by factors such as climate conditions, industrial cycles, and population behavior, necessitate the development of adaptive strategies to maintain a stable energy supply throughout the year. For example, the data from this study indicate that Guangdong and Fujian provinces experience their lowest electricity consumption in July, while Guangxi Zhuang Autonomous Region sees its lowest in January. These variations highlight the need for seasonal energy management practices, such as demand-side management programs, which can shift or reduce energy use during peak periods. Moreover, seasonal tariff adjustments could be introduced to incentivize consumers to reduce their energy usage during high-demand periods, thereby helping to stabilize the energy grid and prevent potential shortages or surpluses. By aligning energy policies with these consumption patterns, policymakers can ensure a more reliable and efficient energy system that is resilient to both short-term fluctuations and long-term trends.

Finally, the insights gained from this study can also contribute to broader policy objectives, such as the transition to a low-carbon economy. Understanding where and when electricity consumption is highest can help in designing policies that promote the use of cleaner energy sources in the most impactful regions and times. For example, areas with high industrial activity, such as Guangdong, could be targeted for the deployment of renewable energy projects, while regions with significant seasonal variability might benefit from energy storage solutions that can balance supply and demand. By leveraging the detailed consumption data provided by NTL analysis, policymakers can not only address immediate energy challenges but also contribute to the long-term sustainability and resilience of the energy system.

4.3. Methodological Limitations and Future Research Directions

While this study provides useful findings and insights, it is important to acknowledge certain methodological limitations that may affect the generalizability and accuracy of the results. One key limitation is the reliance solely on nighttime light (NTL) data as the primary source for estimating electricity consumption. Although NTL data serve as a valuable proxy, they may not fully capture variations in electricity usage influenced by factors such as energy efficiency measures, the adoption of renewable energy, and changes in industrial activity that are not directly reflected in nighttime lighting patterns. In particular, NTL data may introduce selectivity bias, as they are more representative of urban areas with intense lighting, while rural areas or regions with low light emissions may be underrepresented. This bias could result in underestimations of electricity consumption in less populated or less developed areas, where the relationship between light intensity and power usage is not as direct. Additionally, reporting errors could arise from the misclassification of light sources, such as differentiating between industrial, commercial, and residential lighting, which could distort the spatial distribution of estimated electricity consumption.

Furthermore, despite the improved spatial resolution of NPP-VIIRS data compared to previous datasets, challenges like light saturation in highly urbanized areas or inaccuracies in sparsely populated regions with minimal nighttime light emissions could still impact the precision of the estimates. Relying on a single data source such as NTL means that any anomalies in the NTL data can lead to significant deviations in electricity consumption predictions. For instance, by comparing our results with official data from the National Bureau of Statistics of China, we observed that the relatively high electricity consumption estimates for the years 2015–2016 may be due to such anomalies, which could have caused overestimations due to factors like light saturation or misclassification of light sources in densely populated urban areas. Such reporting errors and potential biases need to be accounted for, especially when working with data that aggregate diverse sources of light. Future research should aim to cross-validate NTL-based estimates with other forms of data, such as smart meter readings or direct regional energy usage reports, to mitigate these selectivity biases and improve the overall accuracy of the predictions.

One significant limitation in our analysis arises from the aggregation of pixel-level NTL data to the city level, which introduces the potential for heteroskedasticity in the disturbance term. This occurs because cities vary significantly in terms of their population density, urbanization, and industrial activity. Larger, more industrialized cities with complex economic dynamics might exhibit greater variability in power consumption, leading to more pronounced differences in error variability. We acknowledge that such heteroskedasticity could affect the accuracy of the model’s residuals and predictions. Future research should explore ways to adjust for heteroskedasticity, possibly by using weighted regression techniques or incorporating city-level covariates to better account for these differences.

Another limitation is the use of the gradient boosting regression trees (GBRT) model. Although the GBRT model performed well in this study, it can be susceptible to overfitting, particularly when working with complex and noisy datasets. While cross-validation and regularization techniques were applied to mitigate this risk, future research might benefit from exploring alternative machine learning models, such as deep learning approaches, that may offer improved generalization and adaptability to diverse data conditions. Additionally, integrating other data sources—such as smart meter readings, socioeconomic data, or climate data—could further enhance the accuracy and resolution of electricity consumption estimates. Expanding this methodology to different regions with varying economic and geographical characteristics would also help to test the scalability and adaptability of the approach, providing a more comprehensive understanding of electricity consumption patterns across different contexts.

5. Conclusions

This study provides a comprehensive analysis of electricity consumption in the study area by leveraging nighttime light (NTL) data from NPP-VIIRS and employing advanced machine learning models. Among the models tested, the gradient boosting regression trees (GBRT) model demonstrated the highest accuracy, with an R² value of 0.905, an RD of 13.9%, and an RMSE of 30.6 GWh, making it the most effective model for estimating electricity consumption. The findings reveal significant spatial and temporal variations in electricity consumption, with Guangdong Province consistently showing higher consumption levels due to its robust industrial base, while regions like Guangxi Zhuang Autonomous Region exhibit lower consumption, reflecting their different stages of economic development. This study also underscores the importance of integrating NTL data with other socioeconomic and environmental datasets to enhance the accuracy and resolution of electricity consumption estimates. This approach provides valuable insights for energy planning and policy, particularly in optimizing energy resource allocation, addressing regional disparities, and managing seasonal variations in demand. Overall, the results highlight the potential of using NTL data as a powerful tool for monitoring and understanding regional electricity consumption patterns, contributing to more informed and sustainable energy policies.