A historical data driven method and system for power load forecasting

Abstract

The present invention discloses a historical data driven method and system for power load forecasting, and relates to the technical field of load forecasting. Compared with previous load forecasting models, the present invention solves the problem that an ordinary single power load forecasting model is difficult to adapt to the power load forecasting needs of different durations in different power consumption areas; through the historical data driven method, short-term and long-term power load forecasting of a region or grid can be effectively performed, and the short-term load forecasting model is further optimized by the A3C algorithm. The LSTM network based on the improved particle swarm algorithm for parameter optimization can improve the overall prediction performance. Provide important support for the operation and planning of power systems. The short-term load forecasting model is further adjusted to a long-term load forecasting model based on the improved Monte Carlo simulation method, which can consider a variety of possible scenarios and uncertainty factors, improve the prediction performance of long-term load forecasting, and provide important data support for the efficient operation and optimization of power systems.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting, and in particular to a historical data driven method and system for power load forecasting. Background Art

[0002] With the continuous development and complexity of power systems, accurate power load forecasting has become crucial for the planning, operation and management of power systems. Traditional load forecasting methods mainly rely on manual experience and simple statistical analysis, but these methods often have difficulty in coping with increasingly complex changes in power demand and various uncertainties. The development of machine learning and deep learning algorithms has provided powerful tools for load forecasting. For example, long short-term memory networks (LSTMs) can effectively process long-term dependencies in time series data and capture complex dynamic features in load data.

[0003] However, in different power consumption areas, the power load curve is affected by factors such as seasons, weather conditions, characteristic days and power consumption areas, and presents huge differences in different situations. Therefore, the ordinary single power load forecasting model is difficult to adapt to the power load forecasting needs of different power consumption areas with different durations.

[0004] In order to solve the above problems, the present invention proposes a historical data driven method and system for power load forecasting. Summary of the invention

[0005] The purpose of the present invention is to propose a historical data driven method and system for power load forecasting to solve the problems raised in the background technology:

[0006] The common single power load forecasting model is difficult to adapt to the power load forecasting needs of different power consumption areas and different durations.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A historical data driven method for power load forecasting comprises the following steps:

[0009] S1: Collect historical electricity consumption data in the area and pre-process the data;

[0010] S2: Analyze the preprocessed data and build a short-term load forecasting model based on the A3C algorithm and optimized LSTM network to perform short-term load forecasting;

[0011] S3: Based on the improved Monte Carlo simulation and the short-term load forecasting model built in S2, a long-term load forecasting model is built to perform long-term load forecasting;

[0012] S4: Based on the results of short-term load forecasting and long-term load forecasting, corresponding log records are made, and risk thresholds are divided for risk warning.

[0013] Preferably, in S1, the area is divided into grids, and the electricity consumption data in the corresponding grids are integrated and preprocessed.

[0014] Preferably, the short-term load forecasting model in S2 is built based on an optimized LSTM network, as follows:

[0015] The LSTM network performs feature extraction and preliminary prediction on the historical power load data; the output of the LSTM network is expressed as:

[0016] h t = LSTM ( x t , h t-1 )

[0017] in, x t is the input at the current moment; h t-1 is the hidden state at the previous moment; h t is the hidden state at the current moment;

[0018] The output of LSTM h t Pass it to a fully connected layer to get the predicted load value y t :

[0019] y t = wh t + b

[0020] in, w is the weight parameter; b is the bias parameter, which is obtained by optimizing based on the improved particle swarm algorithm;

[0021] Reinforcement learning of the optimized LSTM network is performed based on the improved A3C algorithm, and the improved A3C algorithm is specifically as follows:

[0022] The reinforcement learning of the improved A3C algorithm is as follows:

[0023] Representing value function based on LSTM network Q ( s c ,a c), indicating that in the state s c Take Action a c The expected return of , using the multi-step Q-learning algorithm to propagate the reward function, as follows:

[0024] Q ( s c ,a c )= R c +φR c+1 +···+ f C-1 R c+C+1 + f C maxQ ( s c+C ,a c )

[0025] in, c , C All are steps; f is the weight factor; R is the reward function;

[0026] The policy network is based on LSTM network learning and optimization strategy to maximize the accuracy of load forecasting. The input of the policy network model is state features, and the output is action.

[0027] Update policy parameters in the network using policy iteration i ; Use the gradient ascent algorithm to predict the expected reward E [ R c ];

[0028] The update of the policy gradient is as follows:

[0029]

[0030] in, π ( a c | s c ; i ) indicates that in the state s c Take Action a c probability; For E [ R c] is an unbiased estimate of ; i is the strategy parameter,

[0031] Strategy parameters i The update method is as follows:

[0032]

[0033] in, β i is a random disturbance; N is the number of sampling samples; α To update the parameters; d is the learning rate parameter;

[0034] The various strategy parameters in the short-term prediction model based on the improved particle swarm algorithm i Optimization;

[0035] v k+1,j = oh k,j + c 1 r 1( p j - u k,j )+ c 2 r 2( g j - u k,j )

[0036] u k+1,j = u k,j +v k+1,j

[0037] in:

[0038] ω=ω min +cos [( k -1) / ( K -1)]·( oh max -oh min )

[0039] c 1= c 11 +[( k -1) / ( K -1)]·( c 11 - c 12 )

[0040] c 2= c 21 +[( k -1) / ( K -1)]·( c 21 - c 22 )

[0041] in, v k,j 、v k+1,j Respectively j The parameter k The second change and k+ 1. The speed of the iteration; u k,j , u k+1,j For the j The parameter k The second change and k+ The value of 1 iteration; oh is the inertia weight; c 1. c 2 is the acceleration factor; r 1. r 2 is a random variable uniformly distributed between [0,1]; p j is the individual optimal value; g j is the global optimal value; K is the maximum number of iterations; oh max 、oh min are the maximum and minimum values ​​of the inertia weight respectively; c 11 , c 12 for c The starting and ending values ​​are 1; c 21 , c 22 for c 2 start and end values.

[0042] Preferably, in S3, the long-term load forecasting model models the volatility of historical load data, and uses a GARCH model to capture the volatility characteristics of load data; based on the GARCH model, a future volatility path is generated, and the future load forecast value is generated by combining the mean of historical load data and the generated volatility path through an improved Monte Carlo simulation, as follows:

[0043]

[0044] in, , for Time and d+ ( l- 1) d The load forecast value at the moment; Ω is the expected value of the predicted load change; x is a constant term; , is the predicted short-term load fluctuation variance; L The number of time divisions, l,e ∈[1, L ]; q l , p e represents a random variable that follows a standard normal distribution; d is the length of the segmentation interval; d To predict the moment; e is a standard normal distribution L A value;

[0045] The future load forecast value generated above is used as an input feature and input into the short-term load forecasting model. The model is trained with historical data, and long-term load forecasting is performed based on the adjusted parameter optimization model.

[0046] A historical data driven system for power load forecasting, comprising:

[0047] Data collection module: used to collect historical electricity consumption data in the area and pre-process the data;

[0048] Short-term load forecasting module: used to analyze the preprocessed data and build a short-term load forecasting model based on the A3C algorithm and optimized LSTM network for short-term load forecasting;

[0049] Long-term load forecasting module: used to build a long-term load forecasting model based on the improved Monte Carlo simulation combined with the short-term load forecasting model built in S2 to perform long-term load forecasting;

[0050] Risk warning module: used to record the corresponding logs based on the results of short-term load forecasting and long-term load forecasting, and divide the risk thresholds for risk warning;

[0051] The data acquisition module includes a grid division unit and a preprocessing unit, which respectively divide the area into grids and integrate and preprocess the electricity consumption data in the corresponding grids.

[0052] Compared with the prior art, the present invention provides a historical data driven method and system for power load forecasting, which has the following beneficial effects:

[0053] The present invention can effectively perform short-term and long-term power load forecasting for a region or grid through a historical data-driven method, and the short-term load forecasting model can improve the overall forecasting performance by further optimizing the LSTM network based on the improved particle swarm algorithm for parameter optimization through the A3C algorithm. It provides important support for the operation and planning of power systems. The short-term load forecasting model is further adjusted to a long-term load forecasting model based on the improved Monte Carlo simulation method, which can consider a variety of possible scenarios and uncertainty factors, improve the forecasting performance of long-term load forecasting, and provide important data support for the efficient operation and optimization of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of the method mentioned in Example 1 of the present invention;

[0055] Figure 2 This is a system block diagram mentioned in Example 2 of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0057] The present invention can effectively perform short-term and long-term power load forecasting for a region or grid through a historical data-driven method, and the short-term load forecasting model is further optimized by the A3C algorithm based on the LSTM network for parameter optimization based on the improved particle swarm algorithm, which can improve the overall forecasting performance. It provides important support for the operation and planning of power systems. The short-term load forecasting model is further adjusted to a long-term load forecasting model based on the improved Monte Carlo simulation method, which can consider a variety of possible scenarios and uncertainty factors, improve the forecasting performance of long-term load forecasting, and provide important data support for the efficient operation and optimization of power systems. Specifically include the following contents.

[0058] Embodiment 1:

[0059] See also Figure 1 The present invention provides a historical data driven method for power load forecasting, comprising the following steps:

[0060] S1: Collect historical electricity consumption data in the area and perform preprocessing operations on the data; the details are as follows:

[0061] Comprehensively collect electricity consumption information for historical regions and grids. This data may come from smart meters, substation records, power company databases, etc. The electricity consumption area is also divided into different grids, and the electricity consumption data in the corresponding grids are integrated. The collected data may contain noise, missing values, and outliers. Data preprocessing is based on data cleaning, outlier processing, missing value filling, etc. For noise, smoothing techniques such as moving average can also be used to reduce its impact. There are various ways to handle missing values, such as filling based on the mean of historical data for the same period, or using more complex interpolation methods, such as spline interpolation. Outliers need to be identified and eliminated by setting reasonable thresholds, or corrected using statistical methods.

[0062] S2: Analyze the preprocessed data and build a short-term load forecasting model based on the A3C algorithm and optimized LSTM network to perform short-term load forecasting; the details are as follows:

[0063] Short-term load forecasting usually predicts load conditions in the next few hours or days and is a key part of power system operation.

[0064] The short-term load forecasting model is built based on the optimized LSTM network, as follows:

[0065] The LSTM network extracts features and makes preliminary predictions on historical power load data; the output of the LSTM network is expressed as:

[0066] h t = LSTM ( x t , h t-1 )

[0067] in, x t is the input at the current moment; h t-1 is the hidden state at the previous moment; h t is the hidden state at the current moment;

[0068] The output of LSTM h t Pass it to a fully connected layer to get the predicted load value y t :

[0069] y t = wh t + b

[0070] in,w is the weight parameter; b is the bias parameter, which is obtained by optimizing based on the improved particle swarm algorithm;

[0071] LSTM can handle long-term dependencies in time series data very well. For example, the electricity consumption information in a residential area has obvious periodicity and trend. LSTM can learn this pattern and accurately predict short-term load changes.

[0072] The reinforcement learning of the LSTM network is optimized based on the improved A3C algorithm. The improved A3C algorithm is as follows:

[0073] The reinforcement learning of the improved A3C algorithm is as follows:

[0074] Representing value function based on LSTM network Q ( s c ,a c ), indicating that in the state s c Take Action a c The expected return of , using the multi-step Q-learning algorithm to propagate the reward function, as follows:

[0075] Q ( s c ,a c )= R c +φR c+1 +···+ f C-1 R c+C+1 + f C maxQ ( s c+C ,a c )

[0076] in, c , C All are steps; f is the weight factor; R is the reward function;

[0077] The policy network is based on LSTM network learning and optimization strategy to maximize the accuracy of load forecasting. The input of the policy network model is state features, and the output is action.

[0078] Update policy parameters in the network using policy iteration i; Use the gradient ascent algorithm to predict the expected reward E [ R c ];

[0079] The update of the policy gradient is as follows:

[0080]

[0081] in, π ( a c | s c ; i ) indicates that in the state s c Take Action a c probability; For E [ R c ] is an unbiased estimate of ; i is the strategy parameter,

[0082] Strategy parameters i The update method is as follows:

[0083]

[0084] in, β i is a random disturbance; N is the number of sampling samples; α To update the parameters; d is the learning rate parameter;

[0085] By continuously adjusting the strategy through the A3C algorithm, the LSTM network can learn a better prediction model, thereby improving the accuracy of power load prediction. The A3C algorithm can be trained simultaneously in multiple threads, which speeds up the training speed and convergence efficiency. Applying it to power load prediction can obtain an effective prediction strategy more quickly.

[0086] Strategy parameters in short-term forecasting model based on improved particle swarm algorithm i Optimization;

[0087] v k+1,j = oh k,j + c 1 r 1( p j - u k,j )+ c 2 r 2( gj - u k,j )

[0088] u k+1,j = u k,j +v k+1,j

[0089] in:

[0090] ω=ω min +cos [( k -1) / ( K -1)]·( oh max -oh min )

[0091] c 1= c 11 +[( k -1) / ( K -1)]·( c 11 - c 12 )

[0092] c 2= c 21 +[( k -1) / ( K -1)]·( c 21 - c 22 )

[0093] in, v k,j 、v k+1,j Respectively j The parameter k The second change and k+ 1. The speed of the iteration; u k,j , u k+1,j For the j The parameter k The second change and k+ The value of 1 iteration; oh is the inertia weight; c 1. c 2 is the acceleration factor; r 1. r 2 is a random variable uniformly distributed between [0,1]; pj is the individual optimal value; g j is the global optimal value; K is the maximum number of iterations; oh max 、oh min are the maximum and minimum values ​​of the inertia weight respectively; c 11 , c 12 for c The starting and ending values ​​are 1; c 21 , c 22 for c The starting and ending values ​​of 2. The improved particle swarm algorithm optimizes the inertia weight and acceleration factor, which can improve the global and local search capabilities of the algorithm and enhance the performance of the prediction model.

[0094] The LSTM network is used as model 1, the parameter-optimized LSTM network is used as model 2, and the optimized LSTM network based on ordinary reinforcement learning is used as model 3. The model performance is evaluated based on multiple evaluation indicators. For details, please refer to Table 1:

[0095] Table 1 Model performance evaluation results of different short-term load forecasting models

[0096]

[0097] S3: Based on the improved Monte Carlo simulation and the short-term load forecasting model built in S2, a long-term load forecasting model is built to perform long-term load forecasting; the details are as follows:

[0098] Long-term load forecasting usually focuses on load change trends in the next few weeks, months or even years, and is crucial for power system planning and resource allocation.

[0099] The long-term load forecasting model models the volatility of historical load data and uses the GARCH model to capture the volatility characteristics of load data. Based on the GARCH model, the future volatility path is generated, and the future load forecast value is generated through the improved Monte Carlo simulation by combining the mean of historical load data and the generated volatility path. The details are as follows:

[0100]

[0101] in, , for Time and d+ ( l- 1) dThe load forecast value at the moment; Ω is the expected value of the predicted load change; x is a constant term; , is the predicted short-term load fluctuation variance; L The number of time divisions, l,e ∈[1, L ]; q l , p e represents a random variable that follows a standard normal distribution; d is the length of the segmentation interval; d To predict the moment; e is a standard normal distribution L A value;

[0102] The future load forecast value generated above is used as an input feature and input into the short-term load forecasting model. The model is trained with historical data, and long-term load forecasting is performed based on the adjusted parameter optimization model.

[0103] A variety of evaluation metrics are used to evaluate the performance of the prediction model, such as mean square error (MSE), mean absolute error (MAE), coefficient of determination (R²), etc.

[0104] According to the evaluation results, the short-term prediction model and the long-term prediction model are adjusted and optimized, such as adjusting the model's hyperparameters, changing the model structure, adding or adjusting input features, etc.

[0105] At the same time, historical data is updated regularly and the model is retrained to adapt to changes in the power system and social economy.

[0106] Similarly, the LSTM network is used as model 1, the parameter-optimized LSTM network is used as model 2, and the optimized LSTM network based on ordinary reinforcement learning is used as model 3. The model performance is evaluated based on multiple evaluation indicators. For details, please refer to Table 2:

[0107] Table 2 Model performance evaluation results of different long-term load forecasting models

[0108]

[0109] As can be seen from Table 1-2, the accuracy of the load model in short-term and long-term load forecasting is higher than that of Model 1, Model 2 and Model 3. In various evaluation indicators, the short-term and long-term load models of this embodiment perform better, which shows that the short-term and long-term load models of this embodiment can more effectively capture the changing laws of power load, thereby providing more accurate and reliable forecasting results.

[0110] S4: Based on the results of short-term load forecasting and long-term load forecasting, corresponding log records are made, and risk thresholds are divided for risk warning. The details are as follows:

[0111] Record the time of each short-term load forecast, the forecast model used, input data characteristics, predicted load values, and actual load values ​​(obtained after the forecast). Record the time, model, input data (such as economic growth data, population changes, industrial planning, etc.), predicted load trends, and specific values ​​of long-term load forecasts.

[0112] The division of short-term risk thresholds and long-term risk thresholds can be made according to the following division method:

[0113] Short-term risk threshold

[0114] High risk: When the predicted load deviates from the actual load by more than 20%, or the actual load exceeds the predicted load upper limit by 15%.

[0115] Medium risk: The deviation is between 10% and 20%, or the actual load exceeds the forecast upper limit by 10% to 15%.

[0116] Low risk: The deviation is between 5% and 10%, or the actual load exceeds the forecast upper limit by 5% to 10%.

[0117] Long-term risk threshold

[0118] High risk: The predicted long-term load growth trend is seriously inconsistent with the actual economic and social development, or the actual load reaches or exceeds the predicted value at the end of the planning period ahead of schedule.

[0119] Medium risk: The long-term load growth rate deviates significantly from expectations (e.g., more or less than 30%).

[0120] Low risk: growth rate deviation is between 10%-30%.

[0121] When the risk threshold is reached, early warning information is sent in a timely manner, including the risk level, the comparison between the forecast and the actual situation, the possible impact and the recommended response measures. For example:

[0122] Short-term high-risk warning: The short-term load forecast deviation exceeds 20%, and the actual load reaches 1,200 kilowatts, while the forecast is 1,000 kilowatts. This may cause local power grid overload. It is recommended to immediately start the backup power supply or perform load regulation.

[0123] Long-term high-risk warning: The long-term load growth trend is far lower than expected. The actual economic slowdown has a significant impact on the growth of electricity demand, which may affect the investment plan of power infrastructure. It is recommended to re-evaluate the feasibility of the project.

[0124] Embodiment 2:

[0125] See also Figure 2 The present invention provides a historical data driven system for power load forecasting, comprising:

[0126] Data collection module 100: used to collect historical electricity consumption data in the area and perform pre-processing operations on the data;

[0127] Short-term load forecasting module 200: used to analyze the preprocessed data and build a short-term load forecasting model based on the A3C algorithm and the optimized LSTM network to perform short-term load forecasting;

[0128] Long-term load forecasting module 300: used to build a long-term load forecasting model based on the improved Monte Carlo simulation combined with the short-term load forecasting model built in S2 to perform long-term load forecasting;

[0129] Model adjustment and optimization module 400: used to make corresponding log records based on the results of short-term load forecasting and long-term load forecasting, and divide risk thresholds for risk warning.

[0130] The data collection module 100 includes a grid division unit 110 and a pre-processing unit 120, which respectively divide the area into grids and integrate and pre-process the power consumption data in the corresponding grids.

[0131] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A historical data driven method for power load forecasting, characterized in that: The steps include: S1: Collect historical electricity consumption data in the area and pre-process the data; S2: Analyze the preprocessed data and build a short-term load forecasting model based on the A3C algorithm and optimized LSTM network to perform short-term load forecasting; the details are as follows: The LSTM network performs feature extraction and preliminary prediction on the historical power load data; the output of the LSTM network is expressed as: h t = LSTM ( x t , h t-1 ) in, x t is the input at the current moment; h t-1 is the hidden state at the previous moment; h t is the hidden state at the current moment; The output of LSTM h t Pass it to a fully connected layer to get the predicted load value y t : y t = wh t + b in, w is the weight parameter; b is the bias parameter, which is obtained by optimizing based on the improved particle swarm algorithm; Reinforcement learning of the optimized LSTM network is performed based on the improved A3C algorithm, and the improved A3C algorithm is specifically as follows: The reinforcement learning of the improved A3C algorithm is as follows: Representing value function based on LSTM network Q ( s c ,a c ), indicating that in the state s c Take Action a c The expected return of , using the multi-step Q-learning algorithm to propagate the reward function, as follows: Q ( s c ,a c )= R c +φR c+1 +···+ φ C-1 R c+C+1 + φ C maxQ ( s c+C ,a c ) in, c , C All are steps; φ is the weight factor; R is the reward function; The policy network is based on LSTM network learning and optimization strategy to maximize the accuracy of load forecasting. The input of the policy network model is state features, and the output is action. Update policy parameters in the network using policy iteration θ ; Use the gradient ascent algorithm to predict the expected reward E [ R c ]; The update of the policy gradient is as follows: in, π ( a c | s c ; θ ) indicates that in the state s c Take Action a c The probability of For E [ R c ] is an unbiased estimate of ; θ is the strategy parameter, Strategy parameters θ The update method is as follows: in, β i is a random disturbance; N is the number of sampling samples; α To update the parameters; δ is the learning rate parameter; The various strategy parameters in the short-term prediction model based on the improved particle swarm algorithm θ Optimization; v k+1,j = ωv k,j + c 1 r 1( p j - u k,j )+ c 2 r 2( g j - u k,j ) u k+1,j = u k,j +v k+1,j in: ω=ω min +cos [( k -1) / ( K -1)]·( ω max -ω min ) c 1= c 11 +[( k -1) / ( K -1)]·( c 11 - c 12 ) c 2= c 21 +[( k -1) / ( K -1)]·( c 21 - c 22 ) in, v k,j 、v k+1,j Respectively j The parameter k The second change and k+ 1. The speed of the iteration; u k,j , u k+1,j For the j The parameter k The second change and k+ The value of 1 iteration; ω is the inertia weight; c 1. c 2 is the acceleration factor; r 1. r 2 is a random variable uniformly distributed between [0,1]; p j is the individual optimal value; g j is the global optimal value; K is the maximum number of iterations; ω max ,ω min are the maximum and minimum values ​​of the inertia weight respectively; c 11 , c 12 for c The starting and ending values ​​are 1; c 21 , c 22 for c 2 start and end values; S3: Based on the improved Monte Carlo simulation and the short-term load forecasting model built in S2, a long-term load forecasting model is built to perform long-term load forecasting; the long-term load forecasting model models the volatility of historical load data and uses the GARCH model to capture the volatility characteristics of load data; based on the GARCH model, a future volatility path is generated, and the future load forecast value is generated through the improved Monte Carlo simulation in combination with the mean of the historical load data and the generated volatility path; the future load forecast value is input as an input feature into the short-term load forecasting model, the model is trained in combination with historical data, and long-term load forecasting is performed based on the adjusted parameter optimization model; S4: Based on the results of short-term load forecasting and long-term load forecasting, corresponding log records are made, and risk thresholds are divided for risk warning.

2. A historical data driven method for power load forecasting according to claim 1, characterized in that: In S1, the area is divided into grids, and the electricity consumption data in the corresponding grids are integrated and preprocessed.

3. A historical data driven method for power load forecasting according to claim 1, characterized in that the generation of the future load forecast value in S3 is specifically as follows: in, , for Time and d+ ( l- 1) d The load forecast value at the moment; Ω is the expected value of the predicted load change; χ is a constant term; , is the predicted short-term load fluctuation variance; L The number of time divisions, l,e ∈[1, L ]; q l , p e represents a random variable that follows a standard normal distribution; d is the length of the segmentation interval; d To predict the moment; ε is a standard normal distribution L A value.

4. A historical data driven system for power load forecasting using the method according to any one of claims 1 to 3, characterized in that: include: Data collection module (100): used to collect historical electricity consumption data in a region and perform pre-processing operations on the data; Short-term load forecasting module (200): used to analyze the preprocessed data and build a short-term load forecasting model based on the A3C algorithm and the optimized LSTM network to perform short-term load forecasting; Long-term load forecasting module (300): used to build a long-term load forecasting model based on improved Monte Carlo simulation combined with the short-term load forecasting model built in S2 to perform long-term load forecasting; Risk warning module (400): used to record corresponding logs based on the results of short-term load forecasting and long-term load forecasting, and divide risk thresholds for risk warning; The data acquisition module (100) comprises a grid division unit (110) and a pre-processing unit (120), which respectively divide the area into grids and integrate and pre-process the electricity consumption data in the corresponding grids.

Images