Power daily load prediction method based on policy quantization and long and short term memory algorithm

Abstract

The invention discloses an electric power daily load prediction method based on policy quantification and a long-short-term memory algorithm, which belongs to the technical field of electric power system load prediction, and comprises the following steps: converting policies into quantitative indexes according to qualitative description of the policies, generating numerical characteristics reflecting the influence of policy changes on electric power demands, and calculating the electric power daily load prediction results; performing data preprocessing through abnormal value cleaning, missing value filling and normalization operation to obtain input data of the model; based on a long short-term memory network, model training is carried out according to input data, and a prediction model is generated by initializing model parameters, setting training hyper-parameters and a self-adaptive training strategy and adjusting dynamic parameters and is used for predicting the daily power load. According to the method, the power load change can be quickly and accurately predicted, so that the uncertainty caused by policy adjustment is effectively dealt with, and the operation efficiency and safety of a power system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting, and in particular to a method for forecasting daily power load based on policy quantification and long short-term memory algorithm. Background Art

[0002] In the current power system, accurate load forecasting is of great significance to the safe and stable operation of the power system and power dispatching. With the continuous adjustment of energy policies, such as carbon emission policies and energy subsidies, traditional power load forecasting methods cannot timely reflect the impact of policy changes on power demand due to the lack of quantitative processing of policy information. Therefore, how to introduce policy changes into the power load forecasting model and combine it with advanced deep learning algorithms, such as long short-term memory networks, to improve the accuracy of load forecasting has become a technical problem that needs to be solved urgently.

[0003] In recent years, with the in-depth development of deep learning theory research, applying deep learning theory to power system electricity demand forecasting has become a very meaningful task. The LSTM model has been widely used in power load forecasting due to its powerful time series data processing capabilities. However, most of the current research relies only on historical data and weather factors, and fails to fully consider the impact of policy changes. Therefore, it is urgent to design a new power daily load forecasting technology that can effectively improve the accuracy of load forecasting. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a daily power load forecasting technology based on policy quantification and long short-term memory algorithm, aiming to effectively deal with the uncertainty caused by policy adjustments and improve the operating efficiency and safety of the power system.

[0005] In order to achieve the above technical objectives, the present application provides a method for daily power load forecasting based on policy quantification and long short-term memory algorithm, comprising the following steps:

[0006] According to the qualitative description of the policy, the policy is converted into quantitative indicators to generate numerical features reflecting the impact of policy changes on electricity demand. The data is preprocessed by cleaning outliers, filling missing values ​​and normalizing operations to obtain the input data of the model.

[0007] Based on the long short-term memory network, the model is trained according to the input data. By initializing the model parameters, setting the training hyperparameters and adaptive training strategies and dynamic parameter adjustments, a prediction model is generated to predict the daily power load.

[0008] Preferably, in the process of obtaining numerical features reflecting the impact of policy changes on electricity demand, the qualitative description of the policy is converted into quantitative indicators based on the historical power load data, meteorological data and policy data in the region to generate numerical features reflecting the impact of policy changes on electricity demand.

[0009] Preferably, in the process of obtaining numerical features, policy indicators, load characteristic indicators, economic indicators, meteorological factor indicators, social factor indicators, market dynamic indicators and historical power load data are used as numerical features, wherein the policy indicators include: energy prices, energy investment, subsidy policies, environmental protection, safety protection, power reform and energy conservation and emission reduction;

[0010] Load characteristic indicators include daily load curve, peak-to-valley difference and load factor;

[0011] Economic indicators: GDP growth rate, industrial output and unemployment rate;

[0012] Meteorological factors indicators: temperature, precipitation, ③ humidity and wind speed;

[0013] Indicators of social factors: population density, family structure and urbanization rate;

[0014] Market dynamics indicators: electricity prices, supply chain efficiency and market demand forecasts;

[0015] Historical power load data, including load over the past 24 hours.

[0016] Preferably, in the process of quantifying the policy, the scoring criteria are used to quantify the policy, including the following steps:

[0017] ① Assign weights to each scoring dimension using the AHP method, and score each policy based on its performance in each dimension. The scores can be divided into three dimensions: amount (0-10 points); coverage (0-10 points); implementation effect (0-10 points). The total score is calculated as:

[0018] Total score = μ1·Amount score + μ2·Coverage score + μ3·Implementation effect score

[0019] Where: μ1, μ2, and μ3 are the weights of the three dimensions respectively;

[0020] ② Calculate the quantitative value: Use binary classification, 1 means implementation, 0 means non-implementation;

[0021] ③ Quantitative index: Quantitative index = ∑(μ i ·l i )

[0022] Where μ i is the weight of different dimensions, li is the corresponding quantized value.

[0023] Preferably, in the process of data preprocessing, the preprocessing process is:

[0024]

[0025] Where: x' i is the processed normative data, x i is a sample of numerical feature i, x min is the minimum sample value of numerical feature i, x max is the maximum sample value of the numerical feature i.

[0026] Preferably, before obtaining input data, a feature engineering method is used to select key features that have the greatest impact on load forecasting and reduce redundant data, which specifically includes the following steps:

[0027] ①L1 regularization: By sparsifying the weights, the most influential features are automatically selected. The loss function of L1 regularization is:

[0028]

[0029] Where: W is the weight of the model, ||W||1 is the L1 norm, λ is the regularization coefficient, y i is the true value of the i samples, f(x i ) is the model input x i The predicted value under ;

[0030] L1 regularization makes the weights of unimportant features tend to 0, thereby selecting important numerical features;

[0031] ②Principal component analysis PCA is used for dimensionality reduction and is expressed as:

[0032] Z=XM

[0033] Among them, X is the normalized feature matrix, M is the PCA transformation matrix, and Z is the feature matrix after dimensionality reduction;

[0034] After obtaining 24 normalized input features, L1 regularization is first used to select the key features for prediction from the 24 features, and then PCA is used to reduce the dimension of the key features to a low-dimensional space. The feature values ​​after dimensionality reduction are used as the final input layer data of the LSTM model, and the number of input layer nodes is also adjusted synchronously.

[0035] Preferably, before model training, weights and biases between layers of the LSTM model are randomly generated according to the number of nodes in the input layer, hidden layer, and output layer of the long short-term memory network, and the normal distribution method is used for parameter initialization.

[0036] Preferably, during the model training process, the hyperparameters of the model are dynamically adjusted in real time through an adaptive optimization algorithm, specifically including: dynamic regularization strength adjustment, dynamic adjustment of gradient clipping threshold, dynamic weight normalization, dynamic adjustment of label smoothing coefficient and dynamic adjustment of sequence inversion frequency.

[0037] Preferably, after the model training is completed, the mean square error MSE, mean absolute percentage error MAPE and determination coefficient R 2 Judging model performance.

[0038] Preferably, a power daily load forecasting system for implementing the power daily load forecasting method comprises:

[0039] The data collection and processing module is used to convert the policy into quantitative indicators based on the qualitative description of the policy, generate numerical features that reflect the impact of policy changes on electricity demand, and perform data preprocessing by cleaning outliers, filling missing values ​​and normalizing operations to obtain the input data of the model;

[0040] The prediction module is used to train the model based on the long short-term memory network according to the input data, generate a prediction model by initializing the model parameters, setting the training hyperparameters and adaptive training strategy and dynamic parameter adjustment, and is used to predict the daily power load.

[0041] The present invention discloses the following technical effects:

[0042] (1) Multi-dimensional data integration: Compared with the use of single-dimensional data, this solution combines historical power load, meteorological data, and policy data. In particular, it quantifies the qualitative description of the policy into quantitative indicators, so that the model can more comprehensively consider the various factors affecting power load and improve the comprehensiveness and accuracy of the forecast.

[0043] (2) Improved prediction accuracy: By using the LSTM neural network to process the time series data of power load, the model can effectively capture long-term dependencies and complex time characteristics. Compared with traditional linear regression models, ARIMA and other methods, it can more accurately predict power load demand, especially in long-term predictions.

[0044] (3) Improve model training efficiency: Through feature selection (L1 regularization) and dimensionality reduction (PCA), the number of redundant features is reduced, the computational complexity of the model is optimized, the training time is shortened, and the operational efficiency of the model under high-dimensional data is improved.

[0045] (4) Enhance the robustness and generalization ability of the model: By dynamically adjusting various hyperparameters such as regularization strength, gradient clipping threshold, weight normalization, label smoothing coefficient, etc., it can adapt to the needs of different training stages and avoid model overfitting or underfitting problems, thereby improving the performance stability of the model in actual application scenarios.

[0046] (5) Improve the adaptability and flexibility of equipment: Dynamically adjusted hyperparameters and adaptive optimization algorithms (such as Adam and RMSprop) can adjust the model's training strategy according to real-time data changes, thereby improving the model's adaptability to different power market demand scenarios and making it more widely applicable.

[0047] (6) Improve work efficiency: The automated feature selection, dynamic parameter adjustment, and adaptive optimization strategies in the solution reduce the need for manual intervention, improve the automation level of the workflow, and improve work efficiency. This enables the prediction system to respond to data changes more quickly, reducing the labor costs required for model adjustment and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 This is the power load forecasting flow chart of the present invention. DETAILED DESCRIPTION

[0050] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0051] like Figure 1 As shown, the present invention provides a method for daily power load forecasting based on policy quantization and long short-term memory (LSTM) algorithm, comprising the following steps:

[0052] (1) Data collection and policy quantification: The historical power load data, meteorological data, and policy data in the specified area are collected, and the qualitative description of the policy is converted into quantitative indicators to generate numerical features reflecting the impact of policy changes on power demand as input data for the model.

[0053] (2) Determine the LSTM model structure: Determine the number of nodes in the input layer, output layer, and hidden layer of the long short-term memory network to ensure that the network structure can handle time series characteristics.

[0054] (3) Data preprocessing: Processing the raw data, including cleaning outliers, filling missing values, and normalization operations, to ensure the quality and consistency of data input.

[0055] (4) Feature selection and dimensionality reduction: Use feature engineering methods to select the key features that have the greatest impact on load forecasting, reduce redundant data, and optimize model efficiency.

[0056] (5) Initialize model parameters: Randomly generate initial parameters, including weights and biases between layers, and set the activation function of the LSTM network to prepare for model training.

[0057] (6) Setting training hyperparameters: In addition to conventional hyperparameters, such as the number of training times, learning rate, batch size, etc., regularization strategy, weight normalization, gradient clipping, label smoothing, and sequence reversal are also used as model training parameters to optimize the training process and improve model performance.

[0058] (7) Adaptive training strategy and dynamic parameter adjustment: Through the adaptive optimization algorithm, the above hyperparameters can be adjusted dynamically in real time.

[0059] (8) Model verification and performance tuning: The model is verified based on adaptive training, with mean square error (MSE), mean absolute percentage error (MAPE), R 2 The coefficient of determination is an indicator for judging the performance of the model.

[0060] (9) After completing the training and optimizing the model, the model is used to forecast the daily power load.

[0061] Specifically, we collect historical power load data, meteorological data, and policy-related data in a specific area, convert the qualitative description of the policy into quantitative indicators, and generate a set of numerical features that can reflect the impact of policy changes on power demand. These numerical features will be used as input data for the model, with a total of 24 input features. The numerical features are divided into the following categories:

[0062] (1) Policy indicators: ① Energy prices, ② Energy investment, ③ Subsidy policy, ④ Environmental protection, ⑤ Safety protection, ⑥ Power reform, ⑦ Energy conservation and emission reduction;

[0063] (2) Load characteristic indicators: ① daily load curve, ② peak-to-valley difference, ③ load factor;

[0064] (3) Economic indicators: ① GDP growth rate, ② industrial output value, ③ unemployment rate;

[0065] (4) Meteorological factors: ① temperature, ② precipitation, ③ humidity, ④ wind speed;

[0066] (5) Social factor indicators: ① population density, ② family structure, ③ urbanization rate;

[0067] (6) Market dynamic indicators: ① electricity price, ② supply chain efficiency, ③ market demand forecast;

[0068] (7) Historical power load data (including the load in the past 24 hours).

[0069] The above policy indicators are quantified using scoring criteria, as shown below:

[0070] ① Assign weights to each scoring dimension using the AHP method, and score each policy based on its performance in each dimension. The scores can be divided into three dimensions: amount (0-10 points); coverage (0-10 points); implementation effect (0-10 points). The total score is calculated as:

[0071] Total score = μ1·Amount score + μ2·Coverage score + μ3·Implementation effect score

[0072] Where: μ1, μ2, and μ3 are the weights of the three dimensions respectively

[0073] ② Calculate the quantitative value: Use binary classification, 1 means implemented, 0 means not implemented.

[0074] ③ Quantitative index: Quantitative index = Σ(μ i ·l i )

[0075] Where μ i is the weight of different dimensions, l i is the corresponding quantized value.

[0076] (2) An LSTM unit has an input gate, a forget gate, and an output gate, which are used to protect and control the unit state. Prepare the model according to the settings in the previous step, and determine the number of nodes h in the input layer of the long short-term memory (LSTM) network (determined by the classification dimension of the input data), the number of nodes l in the output layer (depending on the number of prediction targets), and the number of nodes in the hidden layer.

[0077] The input layer includes 24 input features, including electricity price fluctuations, carbon emission reduction targets, new energy technology R&D investment, electricity market openness, energy conservation and emission reduction incentives, electricity supply and demand gap, historical load data and meteorological data.

[0078] The number of hidden layer nodes can be obtained by:

[0079] Where h is the number of input layer nodes and α is a constant between 1 and 10.

[0080] Since the power load value at a single future time point is predicted, the number of nodes in the output layer is 1.

[0081] (3) Data preprocessing: Process the raw data, including cleaning outliers, filling missing values, and normalizing to ensure the quality and consistency of data input. The unified formula for data preprocessing is:

[0082]

[0083] Where: x' i is the processed normative data, x i is a sample of numerical feature i, x min is the minimum sample value of numerical feature i, x max is the maximum sample value of the numerical feature i.

[0084] (4) After the data is preprocessed, the values ​​of all features are on the same scale. At this time, in order to optimize the model, feature selection and dimensionality reduction are performed.

[0085] ①L1 regularization: By sparsifying the weights, the most influential features are automatically selected. The loss function of L1 regularization is:

[0086]

[0087] Where: W is the weight of the model, ||W||1 is the L1 norm, λ is the regularization coefficient, y i is the true value of the i samples, f(x i ) is the model input x i The predicted value below.

[0088] L1 regularization makes the weights of unimportant features tend to 0, thereby selecting important numerical features.

[0089] ②PCA (Principal Component Analysis): PCA can be used for dimensionality reduction. The formula for PCA dimensionality reduction is:

[0090] Z=XM

[0091] Among them, X is the normalized feature matrix, M is the PCA transformation matrix, and Z is the feature matrix after dimensionality reduction.

[0092] After obtaining 24 normalized input features, L1 regularization is first used to select the most critical features for prediction from the 24 features, and then PCA is used to reduce the dimensions of the most critical features to a low-dimensional space (about 6). The feature values ​​after dimensionality reduction are used as the final input layer data of the LSTM model, and the number of input layer nodes h is also adjusted synchronously.

[0093] (5) According to the determined number of nodes in the input layer, hidden layer, and output layer, the weights and biases between the layers of the LSTM model are randomly generated. For the initial parameters, the normal distribution method is used, and the formula is as follows:

[0094]

[0095] Among them, W (l) represents the weight of the lth layer, with an initial parameter of mean 0 and variance σ 2 The normal distribution of .

[0096] Weight initialization from input layer to hidden layer:

[0097] Each neuron X in the input layer t Connected to each neuron U in the hidden layer i , each connection weight υ ti Random initializations all obey the normal distribution, and the formula is as follows:

[0098]

[0099] Where: t is the dimension of the input feature, and i is the index of the hidden layer neuron.

[0100] Bias initialization of hidden layer neurons:

[0101] Each neuron U in the hidden layer i There is a bias b i , random initializations all obey the normal distribution, the formula is as follows:

[0102] Weight initialization from hidden layer to output layer:

[0103] Each neuron U in the hidden layer i Each neuron Y connected to the output layer j , weight w i Randomly generated, the formula is as follows:

[0104] Bias initialization of output layer neurons:

[0105] Each neuron Y in the output layer j There is a bias b o , random initializations all obey the normal distribution, the formula is as follows:

[0106]

[0107] Select a suitable activation function from Sigmoid, Tanh, ReLU, Leaky Relu and ELU according to the model effect.

[0108] The neuron calculation method of the input layer, hidden layer, and output layer is as follows:

[0109] Hidden layer neurons U i Calculation method:

[0110] Assume X1, X2, X3, ..., X t , ...X n is the input feature, ti The input variable X t With hidden layer neurons U i The weights between the hidden layer neurons U i The output calculation formula is:

[0111]

[0112] Where: f is the activation function of the hidden layer, b i is the neuron U in the hidden layer i The bias.

[0113] Output layer neuron Y j Calculation method:

[0114] The neuron Y in the output layer j Receive output from the hidden layer, assuming that the hidden layer neuron U i By weight w i Connected to the output layer, the calculation formula of the output layer neurons is:

[0115]

[0116] Where: δ is the activation function of the output layer, b o is the bias of the output layer.

[0117] (6) Setting training hyperparameters: In addition to conventional hyperparameters, such as the number of training times, learning rate, batch size, etc., regularization strategy, weight normalization, gradient clipping, label smoothing, and sequence reversal are also used as model training parameters to ensure the convergence, robustness, and generalization ability of the model.

[0118] (7) Adaptive training strategy and dynamic parameter adjustment: Through adaptive optimization algorithms (such as Adam or RMSprop), the above hyperparameters can be adjusted dynamically in real time to improve the training effect and generalization ability of the model.

[0119] ① Dynamic regularization strength adjustment: In the early stage, L1 regularization can help remove unimportant features. In the later stage, L2 regularization is used to limit the size of the model weight to prevent overfitting. Therefore, it is necessary to dynamically adjust the regularization coefficients λ1 and λ2 at different stages of training. The formula is as follows:

[0120]

[0121] Where: ini is the initial regularization coefficient, T half For the later stage of training, fin is the regularization strength in the later stage of training.

[0122] ② Dynamic adjustment of gradient clipping threshold: In the early and late stages of training, the gradient fluctuations of the model may be different, so the gradient clipping threshold needs to be adjusted dynamically. The formula is as follows:

[0123]

[0124] Where: g ini is the initial gradient clipping threshold. As the number of training times t increases, the clipping threshold gradually decreases, and the gradient update in the model convergence stage becomes more stable.

[0125] ③ Dynamic weight normalization: When the weight update is large in the early stage, a weaker normalization is performed. In the later stage of training, the error continues to decrease and the normalization strength is strengthened. Therefore, in order to stabilize the model, it is necessary to dynamically adjust the weight normalization strength. The formula is as follows:

[0126]

[0127] Where: W nor，t represents the standardized weight in the tth training, W nor，t-1 represents the standardized weight in the t-1th training, and η is the learning rate adjustment coefficient for weight standardization.

[0128] ④ Dynamic adjustment of label smoothing coefficient: In the early stage of training, in order to avoid overfitting, a higher label smoothing coefficient is used. In the later stage of training, weaker label smoothing allows the model to focus more on learning the current task. The formula is as follows:

[0129]

[0130] Where: ∈ t is the label smoothing coefficient in the t-th round of training, ∈ iniis the initial label smoothing coefficient, t is the current training round, T max is the maximum number of training rounds.

[0131] ⑤ Dynamic adjustment of sequence reversal frequency: In the early stage of training, in order to help the model capture time features, the model performs sequence reversal with a higher probability. In the later stage of training, the reversal probability is reduced so that the model can focus more on learning the natural time sequence in the later stage of training. The formula is as follows:

[0132]

[0133] Where: p rev，t is the probability of sequence reversal in the tth round of training, p ini is the probability of initial sequence reversal.

[0134] (8) Model verification and performance tuning: The model is verified based on adaptive training, with mean square error (MSE), mean absolute percentage error (MAPE), R 2 The coefficient of determination is an indicator for judging the performance of the model. The formula is as follows:

[0135]

[0136] Where: y i is the true value of the i-th sample, is the predicted value of the i-th sample. The smaller the MSE, the higher the prediction accuracy.

[0137] ② The smaller the MAPE, the smaller the relative error of the model prediction.

[0138] ③

[0139] Where: is the average value of the true value, R 2 ∈[0, 1], R 2 The closer it is to 1, the better the model fits the data.

[0140] (9) After completing the training and optimizing the model, the model is used to forecast the daily power load.

[0141] In response to the problem that power load is greatly affected by policy changes (such as carbon emission policies, energy subsidies, etc.), the present invention proposes a long short-term memory algorithm based on policy quantification and improvement, and establishes a prediction model that reflects the nonlinear relationship between policy changes and power load. It can help power dispatchers quickly and accurately predict power load changes, thereby effectively coping with the uncertainty brought about by policy adjustments and improving the operating efficiency and safety of the power system.

[0142] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0143] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0144] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting daily power load based on policy quantification and long short-term memory algorithm, characterized in that: The following steps are involved: According to the qualitative description of the policy, the policy is converted into quantitative indicators to generate numerical features reflecting the impact of policy changes on electricity demand, and data preprocessing is performed by cleaning outliers, filling missing values ​​and normalizing operations to obtain the input data of the model; Based on the long short-term memory network, model training is performed according to the input data, and a prediction model is generated by initializing model parameters, setting training hyperparameters and adaptive training strategies and dynamic parameter adjustments to predict daily power load.

2. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 1 is characterized by: In the process of obtaining numerical features that reflect the impact of policy changes on electricity demand, the qualitative description of the policy is converted into quantitative indicators based on the historical power load data, meteorological data and policy data in the region to generate numerical features that reflect the impact of policy changes on electricity demand.

3. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 2 is characterized by: In the process of obtaining the numerical features, policy indicators, load characteristic indicators, economic indicators, meteorological factor indicators, social factor indicators, market dynamic indicators and historical power load data are used as the numerical features, wherein the policy indicators include: energy prices, energy investment, subsidy policies, environmental protection, safety protection, power reform and energy conservation and emission reduction; The load characteristic indicators include daily load curve, peak-to-valley difference and load factor; The economic indicators mentioned: GDP growth rate, industrial output value and unemployment rate; The meteorological factors indicators: temperature, precipitation, ③ humidity and wind speed; The social factors indicators mentioned are: population density, family structure and urbanization rate; The market dynamic indicators mentioned: electricity prices, supply chain efficiency and market demand forecasts; The historical power load data includes the load in the past 24 hours.

4. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 3 is characterized by: In the process of quantifying policies, scoring criteria are used to quantify policies, including the following steps: ① Assign weights to each scoring dimension using the AHP method, and score each policy based on its performance in each dimension. The scores can be divided into three dimensions: amount (0-10 points); coverage (0-10 points); implementation effect (0-10 points). The total score is calculated as: Total score = μ1·Amount score + μ2·Coverage score + μ3·Implementation effect score Where: μ1, μ2, and μ3 are the weights of the three dimensions respectively; ② Calculate the quantitative value: Use binary classification, 1 means implementation, 0 means non-implementation; ③ Quantitative index: Quantitative index = ∑(μ i ·l i ) Where μ i is the weight of different dimensions, l i is the corresponding quantized value.

5. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 4 is characterized by: In the process of data preprocessing, the preprocessing process is: Where: x' i is the processed normative data, x i is a sample of numerical feature i, x min is the minimum sample value of numerical feature i, x max is the maximum sample value of the numerical feature i.

6. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 5 is characterized by: Before obtaining input data, feature engineering methods are used to select key features that have the greatest impact on load forecasting and reduce redundant data. The specific steps include: ①L1 regularization: By sparsifying the weights, the most influential features are automatically selected. The loss function of L1 regularization is: Where: W is the weight of the model, ||W||1 is the L1 norm, λ is the regularization coefficient, y i is the true value of the i samples, f(x i ) is the model input x i The predicted value under ; L1 regularization makes the weights of unimportant features tend to 0, thereby selecting important numerical features; ②Principal component analysis PCA is used for dimensionality reduction and is expressed as: Z=XM Among them, X is the normalized feature matrix, M is the PCA transformation matrix, and Z is the feature matrix after dimensionality reduction; After obtaining 24 normalized input features, L1 regularization is first used to select the key features for prediction from the 24 features, and then PCA is used to reduce the dimension of the key features to a low-dimensional space. The feature values ​​after dimensionality reduction are used as the final input layer data of the LSTM model, and the number of input layer nodes is also adjusted synchronously.

7. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 6 is characterized by: Before model training, the weights and biases between the layers of the LSTM model are randomly generated according to the number of nodes in the input layer, hidden layer, and output layer of the long short-term memory network, and the normal distribution method is used to initialize the parameters.

8. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 7 is characterized by: During the model training process, the model's hyperparameters are adjusted dynamically in real time through an adaptive optimization algorithm, including: dynamic regularization intensity adjustment, dynamic adjustment of gradient clipping threshold, dynamic weight normalization, dynamic adjustment of label smoothing coefficient, and dynamic adjustment of sequence inversion frequency.

9. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to claim 8 is characterized by: After the model training is completed, the mean square error MSE, mean absolute percentage error MAPE and determination coefficient R 2 Judging model performance.

10. The method for daily power load forecasting based on policy quantification and long short-term memory algorithm according to any one of claims 1 to 9, characterized in that: The electric power daily load forecasting system for implementing the electric power daily load forecasting method comprises: The data acquisition and processing module is used to convert the policy into quantitative indicators based on the qualitative description of the policy, generate numerical features reflecting the impact of policy changes on electricity demand, and perform data preprocessing by cleaning outliers, filling missing values ​​and normalizing operations to obtain the input data of the model; The prediction module is used to perform model training based on the long short-term memory network according to the input data, generate a prediction model by initializing model parameters, setting training hyperparameters and adaptive training strategies and dynamic parameter adjustments, and is used to predict the daily power load.

Images