A land use based electricity consumption characterization and prediction method and system

Abstract

The application belongs to the technical field of power consumption prediction, and relates to a land use-based power consumption representation and prediction method and system, which comprises the following modules: a data preprocessing module, a land function identification / land use fine processing module, a land use-power consumption explainable machine learning model construction module, a future land use prediction module, and a power representation / prediction and strategy proposal module; the data set generated by the first module is respectively sent to the second, third, fourth and fifth modules as input data of the modules according to the requirements of the modules; the second, third and fourth modules respectively send the generated data to the fifth module as input data of the fifth module, which is finally used for assisting in strategy making; compared with the prior art, the application is easier to carry out and has better explainability.

Description

Technical Field

[0001] This invention belongs to the field of electricity consumption prediction and environmental modeling, and relates to a method and system for characterizing and predicting electricity consumption based on land use. Background Technology

[0002] The characterization and forecasting of electricity consumption plays a positive role in regional power generation, electricity pricing, renewable energy integration, and energy system optimization and management. Therefore, conducting multi-scale (national, provincial, municipal, district / county, township / street, industrial park) electricity consumption forecasting has significant theoretical and practical implications. Currently, methods for characterizing and forecasting electricity consumption include the following:

[0003] 1) Socioeconomic indicators: Based on economic and population statistics from statistical yearbooks, establish the relationship between electricity consumption and the above indicators (including BPNN, etc.), and perform scenario pre-setting and forecasting. In addition, there are time series forecasting methods such as LSTM.

[0004] 2) Nighttime light data: Based on the DMSP / OLS and NPP / VIIRS nighttime light datasets, a machine learning model was used to establish the relationship between nighttime light and electricity consumption. Subsequently, nighttime light forecasting was performed through time series prediction, thereby predicting electricity consumption.

[0005] However, the first method suffers from complex autocorrelation and multicollinearity among socioeconomic indicators, making it difficult to guarantee the independence of independent variables during model construction. This results in insufficient generalization ability and interpretability. Furthermore, socioeconomic indicators themselves are difficult to adjust, offering insufficient guidance for specific strategy formulation. In contrast, the land use type at each location in the study area is unique, and the variables are theoretically independent. Moreover, the spatial representation of regional planning is land use planning, allowing the prediction results to be directly used for decision-making and strategy formulation. Introducing an interpretable machine learning model further addresses the poor interpretability of traditional methods.

[0006] Regarding the second method: nighttime light data has a short time series and low spatial resolution (DMSP / OLS: 1992 to 2013, about 20 years, but quite old, with a spatial resolution of 1 km; NPP / VIIRS: 2014 to present, less than 10 years, with a spatial resolution of about 500 meters). Although many studies have attempted to unify the scale of the two data products through machine learning, it has been difficult to achieve so far. Therefore, this method suffers from the problems of limited existing data and poor spatial resolution. Land use data (LULC), combined with spatiotemporal big data such as POIs, can achieve refined land use processing for nearly 13 years since 2012, and can construct data with more spatial features, high spatial resolution (about 30m), recent data, and a longer time series. Furthermore, nighttime light data has continuous variables in each pixel, making prediction difficult, while land use is a categorical variable and can be predicted using methods such as cellular automata and Markov chains.

[0007] Therefore, there is a need for a method or apparatus for characterizing and predicting land use and electricity consumption that is easy to conduct and has good interpretability to solve this technical problem. Summary of the Invention

[0008] The technical solution adopted by the present invention to solve the technical problem is: a method and system for characterizing and predicting electricity consumption based on land use, comprising: Module 1: data preprocessing module; Module 2: land function identification / land use refinement processing module; Module 3: land use-electricity consumption interpretability machine learning model construction module; Module 4: future land use prediction module; Module 5: electricity characterization / prediction and strategy proposal module.

[0009] Module 1 is used to preprocess the initial data as needed and generate a dataset for use by subsequent modules; the initial data includes: multivariate spatiotemporal big data, satellite remote sensing imagery, and statistical yearbook panel data.

[0010] Module 2 is used to refine the land use type of "construction land" in traditional land use data based on multi-dimensional spatiotemporal big data. Traditional land use data includes six categories: cultivated land, forest land, grassland land, water land, construction land, and unused land.

[0011] Module 3 is used to supplement the electricity consumption data obtained by interpolation in Module 1 and the refined land use data obtained in Module 2. A machine learning algorithm is selected for pre-training, the model with the best results is selected, hyperparameter optimization is performed using grid search, and finally the model is trained. For the trained model, an interpretable machine learning model is used to reveal the marginal impact curve of land use on electricity consumption, and a land use-electricity consumption regression model is obtained.

[0012] Module 4 is used to predict large-scale, multi-class land use through optimization algorithms, thereby enabling future land use forecasting; Module 4 is also used to reveal the driving factors of land use expansion and their contribution.

[0013] Module 5 is used to achieve clustering and spatial feature identification of spatial heating, cooling, and electricity based on prediction / characterization results using spatial econometrics, and to assist in strategy formulation based on the renewable energy suitability atlas determined in Module 1; Module 5 is also used to perform dynamic characterization of electricity consumption based on real-time data from Module 1 and Module 2; Module 5 is also used to provide optimal land use planning schemes based on the curves obtained in Module 3, and to ensure the effective implementation of the planning schemes based on the conclusions of Module 4.

[0014] Module 1 generates datasets and sends them to Modules 2, 3, 4, and 5 as input data according to the needs of each module. Modules 2, 3, and 4 then send their generated data to Module 5 as input data, ultimately used to assist in strategy formulation. The model in this invention is not merely for electricity prediction. Modules 1, 2, and 3 respectively: Module 1: Environmental modeling preparation—providing a large amount of feature information for the study area to enable subsequent simulations; Module 2: Refining land use, which is also environmental modeling and a process; Module 3: Interpretable machine learning model, which also reveals the relationship between environmental factors and electricity consumption. Therefore, the model in this invention is not just for prediction, but also a global information model.

[0015] Preferably, the dataset generated after preprocessing in Module 1 includes: a meteorological parameter dataset, a socio-economic panel dataset, a land use planning geographic dataset, a long-term series vector POI geographic dataset, a built-up area street block vector dataset, a spatial parameter dataset, and a long-term series electricity consumption dataset for prefecture-level cities / study areas. After the meteorological parameter dataset, socio-economic panel dataset, land use planning geographic dataset, long-term series vector POI geographic dataset, and spatial parameter dataset are input data to the data input terminal of Module 2, Module 2 generates a global-scale refined land use processing result. After the long-term series vector POI geographic dataset is input data to the data input terminal of Module 4, Module 4 generates a future land use prediction result under the current development scenario. The long-term series electricity consumption dataset for prefecture-level cities / study areas is input data to Module 3. The meteorological parameter dataset, the global-scale refined land use processing result, and the future land use prediction result under the current development scenario are input data to Module 5 for assisting strategy formulation.

[0016] Preferably, the preprocessing steps of the initial data in Module 1 include: when processing meteorological and geothermal test parameters, using geographic interpolation to obtain a climate, meteorological, and geothermal raster dataset covering the study area; when processing socioeconomic indicators, if data for year n is missing, but data for years n-1 and n+1 exists, then linear interpolation is used; otherwise, ARIMA interpolation is used; after extracting land use identification data, geographic links are generated to make the results include geographic coordinates; POI data and place name feature data are obtained using Python; statistical yearbook data uses the same interpolation method as socioeconomic indicators.

[0017] Preferably, the refinement process of module two includes: calculating the kernel density of the preprocessed POI data, using the kernel density result as the independent variable X of the training dataset, using the labeled land use data as the dependent variable Y of the training dataset, and using the intersection of the two as training samples; using an artificial neural network or random forest algorithm as a regression model for data training; using the kernel density data of unlabeled land use data team members as independent variables, inputting them into the trained model to obtain the global dependent variable, and finally obtaining the refined result of land use.

[0018] Preferably, the land use prediction steps in Module 4 include: using past socioeconomic raster data and natural environment raster data obtained in the data preparation stage as independent variables X; using the land use refinement results obtained in Module 2 as Y according to the time-varying raster data, and training a land expansion prediction model; using a variable search step size when selecting training samples, sampling every k units when encountering ecological land use, and sampling with a single step size when encountering built-up areas, where the threshold k is obtained through a grid irrelevance test; importing the sampling results into a random forest model for training to obtain the relationship between X and Y; and then, calculating a probability atlas of future land use expansion based on the latest socioeconomic raster data and natural environment raster data.

[0019] Preferably, the steps for implementing multi-scenario simulation and regional development pre-setting in Module 5 include: predicting future land use based on the current refined land use results and the expanded probability atlas obtained in Module 4; using future land use as the independent variable X, selecting the model trained in Module 3 to predict future electricity consumption and its spatial distribution; conducting spatial econometric analysis on the prediction results to identify hot and cold areas, clusters, and specific values, and selectively applying renewable energy in these areas based on the renewable energy evaluation results obtained in Module 1, and planning the optimal land use layout based on the interpretability of the model in Module 3.

[0020] More preferably, in module five, when predicting land use over a smaller area, the prediction models include: PLUS model, FLUS model, and CLUE-S model.

[0021] Preferably, when selecting a machine learning algorithm for pre-training in module three, the machine learning algorithms include: multiple linear regression, BP neural network, MLP, random forest, XG Boost, Extra Tree, OLS, and ANN; when using an interpretable machine learning model to establish the relationship between variables, the machine learning models include: SHapley Additive exPlanation and SHAP model.

[0022] The beneficial effects of this invention are:

[0023] 1. This invention addresses the problem that satellite remote sensing image interpretation cannot reflect the specific functions of construction land through Module 1 and Module 2: Land use data is acquired through satellite remote sensing image interpretation, which relies on the wavebands reflected from the land surface and therefore cannot identify the specific functions of construction land. POI vector data often carries functional labels, but POI data from previous years is difficult to obtain, and crawling parameters such as segmentation thresholds are inconsistent. Spatiotemporal big data such as review data have year information labels; therefore, by extracting keywords from review data using natural language processing / semantic segmentation and establishing links with POIs, long-term, uniform-scale POI data can be obtained. Based on this, Chinese land use types are divided according to street blocks to determine the functions of each block, and the aforementioned POIs are applied to achieve refined land use processing.

[0024] 2. This invention establishes a highly interpretable regression model for land use and electricity consumption based on game theory in Module 3. Traditional multiple linear regression has strong interpretability but low goodness of fit, making it difficult to use for prediction. Machine learning models, on the other hand, have high goodness of fit and strong learning ability for data features, but weak interpretability. Therefore, it is necessary to interpret the machine learning model to provide a reference for strategy and to obtain the marginal impact of different independent variables on the dependent variable. Therefore, by using sample replacement / illness data attack on the trained model, the marginal impact and combined effect of each independent variable on the dependent variable under different values ​​can be obtained.

[0025] 3. This invention achieves large-scale, multi-class land use prediction through Module Four: Large-scale, multi-class land use prediction requires substantial computational resources, while arable land, forest land, grassland, water bodies, and unused land have relatively low impacts on electricity consumption. Therefore, sacrificing the prediction accuracy of these land uses for higher efficiency is feasible. These land uses often appear in clusters. Reasonable thresholds are determined based on the normal / skewed distribution curves of the data. When arranging the prediction grid, a reasonable grid distribution is selected based on the data distribution, and the impact of the grid on the simulation results is evaluated / grid inconsistency (or weak correlation) is verified. Based on this, random seeds are set according to the grid layout, and development thresholds for different regions are determined according to the development scenario, ultimately achieving large-scale, multi-class land use simulation and prediction. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the inter-module connections of a method and system for characterizing and predicting electricity consumption based on land use.

[0027] Figure 2 This is a diagram illustrating the data input / output connection relationships between modules;

[0028] Figure 3 This is the flowchart for module one;

[0029] Figure 4 This is the flowchart for Module 2;

[0030] Figure 5 This is the flowchart for module three;

[0031] Figure 6 This is the flowchart for module four;

[0032] Figure 7 This is the flowchart for Module 5. Detailed Implementation

[0033] The related technologies of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] refer to Figures 1-7 , Figure 1 This is the general technical approach of this embodiment. Figures 3 to 7 This is a flowchart of the processing for each module. This implementation method is divided into 5 main modules:

[0035] Module 1: Data Preprocessing Module. Preprocessing is performed on multivariate spatiotemporal big data, satellite remote sensing imagery, and statistical yearbook panel data as needed. The purpose of multivariate spatiotemporal big data preprocessing is to achieve four main objectives (serving the evaluation of renewable energy proposed in subsequent strategies, serving the refined processing of land use, serving the independent and dependent variables and constructing correlations, and serving land use prediction), as detailed below:

[0036] When processing meteorological and geothermal test parameters, geographic interpolation (i.e., Kriging interpolation in ArcGIS) is used to obtain a climate, meteorological, and geothermal raster dataset covering the study area. When processing socioeconomic indicators, if data for year n is missing but data for years n-1 and n+1 exists, linear interpolation is used; otherwise, ARIMA interpolation is used. Land use identification data is extracted from the "xxx City Territorial Spatial Planning (20xx-20xx)" published on the website, and geographic links are generated to include geographic coordinates in the results. POI data and place name feature data are obtained using Python from Gaode Maps and Meituan / Dianping / Baidu, respectively. Road network data comes from the OSM platform; dead-end roads may exist for road elements, requiring cover processing. Statistical yearbook data uses the same interpolation method as the "socioeconomic indicators."

[0037] 2) Module Two: Refined Land Use Processing. This module, based on traditional land use data (6 categories: cultivated land, forest land, grassland land, water land, construction land, and unused land), uses multi-dimensional spatiotemporal big data to refine the land use type of "construction land" within the 6 categories, as detailed below:

[0038] Preprocessed POI data is used to calculate kernel density using ArcGIS. The kernel density result is used as the independent variable (X) of the training dataset, and the labeled land use data is used as the dependent variable (Y) of the training dataset. The intersection of the two is used as the training sample. An artificial neural network or random forest algorithm is used as the regression model for data training. Kernel density data of unlabeled land use data members is used as the independent variable and input into the trained model to obtain the global dependent variable (i.e., the refined land use result). Finally, the refined land use result is obtained.

[0039] 3) Module Three: Based on the interpolated and imputed electricity consumption data from Module One (as the dependent variable Y) and the refined land use data obtained from Module Two (as the independent variable X), machine learning algorithms such as OLS, MLP, ANN, XGBoost, and Extra Tree are selected for pre-training. The model with the best results (best R-squared on the test set) is selected, and hyperparameter optimization is performed using a grid search method. Finally, the model is trained. For the trained model, an interpretable machine learning model (such as the SHapley Additive exPlanation, SHAP model) is used to establish the relationships between variables. This yields a land use-electricity consumption regression model.

[0040] Module 4: Conducting Future Land Use Prediction. This module enables large-scale, multi-class land use prediction through algorithm optimization. Details are as follows:

[0041] The historical socioeconomic and natural environment raster data obtained during the data preparation phase are used as independent variables (X). The land use refinement results obtained in Module 2 are used as Y, with the raster representing changes over time, to train a land expansion prediction model. A variable search step size (a self-proposed optimization algorithm) is used when selecting training samples. When encountering ecological land use, sampling is performed every k cells; when encountering built-up areas, sampling is performed with a single step size. The threshold k is obtained through a grid independence test (gradually increasing the density of the search grid / random seed number, selecting small areas, evaluating the convergence curve of the model accuracy, and obtaining the optimal k value). The sampling results (including X and Y) are imported into a random forest model for training to obtain the relationship between X and Y (referencing the LEAS model approach). Furthermore, based on the latest socioeconomic and natural environment raster data, a probability atlas of future land use expansion is calculated. (Response: This is a CFD-based heuristic algorithm optimization, and there are currently no relevant examples. The underlying principle is that for some coarse land use, it is not necessary to draw very dense grids for simulation, while in the GIS field, grids of uniform size are generally used. The question is how to densify the grid locally, and by how much, using grid independence verification—that is, gradually increasing the density until the impact on the result is not significant, which is the threshold.)

[0042] 5) Module Five: Based on the prediction / representation results, spatial econometrics is used to achieve clustering and spatial feature identification such as spatial heating, cooling, and electricity. The renewable energy suitability atlas determined in Module One assists in strategy formulation. Furthermore, this module can establish a correlation between socio-economic factors and land use, enabling multi-scenario simulation and regional development pre-planning. Details are as follows:

[0043] Based on the current refined land use results and the expanded probability atlas obtained in Module 4, future land use is predicted. Using future land use as the independent variable (X), the model trained in Module 3 is selected to predict future electricity consumption and its spatial distribution. Spatial econometric analysis is performed on the prediction results to identify hot and cold areas, clusters, and specific values. In these areas, renewable energy applications (technology strategies) are selected based on the renewable energy evaluation results obtained in Module 1, and planning is carried out based on the optimal land use layout obtained from the interpretability of the model in Module 3 (planning and management strategies).

[0044] In this embodiment, "Conclusion 1" and "Conclusion 2" provided in Modules 3 and 4 will be used to construct a global information model in the characterization phase, and are also an important part of Module 5. In addition, the data preprocessing of each element in Module 1 is also an important component of environmental modeling; prediction is one part, and the information model / development model is also an important part of this invention.

[0045] In summary, this invention makes electricity consumption characterization and prediction easier to conduct and provides better interpretability by using land use for electricity consumption characterization and prediction, refining land use, and simulating large-scale multi-class land use.

[0046] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A land use based electricity consumption characterization and forecasting method, characterized in that, It comprises: module one: data preprocessing module, module two: land function identification / land use fine processing module, module three: land use-power consumption explainable machine learning model construction module, module four: future land use prediction module, module five: power characterization / prediction and strategy proposal module. The module one is used for preprocessing initial data and generating a data set for subsequent modules according to needs; the initial data includes: multi-temporal and spatial big data, satellite remote sensing images, and statistical yearbook panel data. The module two is used for fine processing of the land use type of "construction land" in traditional land use data based on multi-temporal and spatial big data; the traditional land use data includes: six categories of cultivated land, forest land, grassland, water, construction land, and unused land. The module three is used for selecting a machine learning algorithm for pre-training based on the power consumption data interpolated and supplemented by the module one and the fine land use data obtained by the module two, selecting the best model, adopting a grid search method for hyperparameter optimization, and finally training; for the trained model, an explainable machine learning model is used to reveal the marginal influence curve of land use on power consumption, and a land use-power consumption regression relationship model is obtained. The module four is used for predicting large-scale multi-classification land use through an optimization algorithm, so as to carry out future land use prediction; the module four is also used for revealing the expansion driving force factors of land use and their contribution degree. The module five is used for realizing clustering, spatial cold-heat power spatial feature identification based on spatial econometrics based on the prediction / characterization results, and assisting strategy formulation based on the renewable energy suitability map set determined by the module one; the module five is also used for dynamically characterizing power consumption based on real-time module one and module two data; the module five is also used for providing an optimal land use planning scheme based on the curve obtained by the module three, and ensuring effective implementation of the planning scheme based on the conclusion of the module four. The module one generates a data set and sends it to the module two, the module three, the module four, and the module five as input data of each module according to the needs of each module; the module two, the module three, and the module four send the generated data to the module five as input data of the module five for assisting strategy formulation. The steps of realizing multi-scenario simulation and regional development preset in the module five include: predicting future land use based on the current land use fine results and the expansion probability map set obtained by the module four; taking the future land use as an independent variable X, selecting the model trained by the module three to predict future power consumption and the spatial distribution of power consumption; performing spatial econometrics analysis on the prediction results to identify cold and hot spot regions, clustering areas, specific values, and selecting renewable energy applications in these areas according to the renewable energy evaluation results obtained by the module one, and planning based on the optimal land use layout obtained by the explainability of the module three model. In the fifth module, for the prediction of land use in a smaller range, the prediction model comprises: PLUS model, FLUS model, CLUE-S model; In the third module, when the machine learning algorithm is selected for pre-training, the machine learning algorithm comprises: multiple linear regression, BP neural network, MLP, random forest, XG Boost, Extra Tree, OLS, ANN; when the interpretable machine learning model is used to establish the relationship between variables, the machine learning model comprises: SHapley Additive exPlanation, SHAP model.

2. A land use based electricity consumption characterization and forecasting method as claimed in claim 1, wherein, The data set generated after preprocessing in the first module comprises: meteorological parameter data set, social and economic panel data set, land use planning geographic data set, long time series vector POI geographic data set, built-up area block vector data set, spatial parameter data set, and long time series prefecture-level city / research area power consumption data set; after the meteorological parameter data set, the social and economic panel data set, the land use planning geographic data set, the long time series vector POI geographic data set, and the spatial parameter data set are input as input data to the data input end of the second module, the second module generates a global range land use refinement result; after the long time series vector POI geographic data set is input as input data to the data input end of the fourth module, the fourth module generates a future land use prediction result under an existing development scenario; the long time series prefecture-level city / research area power consumption data set is input as input data to the third module; the meteorological parameter data set, the global range land use refinement result, and the future land use prediction result under the existing development scenario are input as input data to the fifth module for assisting in strategy making.

3. The method of claim 1, wherein, The preprocessing step of the initial data in the first module comprises: when processing meteorological and geothermal test parameters, using geographic interpolation method to obtain climate meteorological and geothermal raster data set covering the research area; when processing social and economic indicators, if the data of the nth year is missing, and the data of the n-1th year and the n+1th year exists, then linear interpolation method is used, otherwise ARIMA interpolation method is used; after the identification data of land use is extracted, the result is linked with geographic coordinates; POI data and place name feature data are obtained through Python; statistical yearbook data uses the same interpolation method as the social and economic indicators.

4. The method of claim 1, wherein, The refinement step of the second module comprises: calculating the kernel density of the preprocessed POI data, taking the kernel density result as the independent variable X of the training data set, taking the identified land use data as the dependent variable Y of the training data set, and taking the intersection of the two as the training sample; using artificial neural network or random forest algorithm as the regression model for data training; taking the kernel density data of the land use data without identification as the independent variable, inputting it into the trained model to obtain the global dependent variable, and finally obtaining the refinement result of land use.

5. The method of claim 1, wherein, The step of land use prediction in the fourth module comprises: taking the past social and economic grid data and the natural environment grid data obtained in the data preparation stage as independent variables X; taking the land use refinement result obtained in the second module as Y, and establishing a land expansion prediction model training according to the change of the grid over time; when selecting the training samples, a variable search step is adopted, when encountering ecological land, the next sampling is performed every k cells, and when encountering the built-up area, a single step sampling is adopted, wherein the threshold value k is obtained through the grid independence test; the sampling result is introduced into the random forest model for training to obtain the relationship between X and Y; and then, based on the latest social and economic grid data and the natural environment grid data, a future land use expansion probability atlas is calculated.

Images