Short-term net load prediction method

Abstract

The invention relates to a short-term net load prediction method, which comprises a data preprocessing technology based on historical net load data, decomposed net load data, weather data, date data and other factors, and the data preprocessing technology comprises data cleaning, feature selection and feature scaling. Wherein the data cleaning is used for processing error, missing or invalid data in the data set, and comprises the steps of deleting repeated data and processing missing data and abnormal values. The method has the beneficial effects that 1, the prediction accuracy is improved; 2, the generalization ability of the model is improved; and 3, the adaptability and flexibility of the model are enhanced. And 4, the data utilization efficiency is improved. And 5, efficient operation and management of a power system are facilitated. In an electricity market environment, accurate net load prediction can help an electricity supplier and a user to formulate a reasonable transaction strategy, the market risk is reduced, the cost is controlled, and the healthy development of the electricity market is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of parameter perception and fault handling methods of verification systems, and in particular relates to a short-term net load prediction method. Background Art

[0002] Short-term net load forecasting is of great significance in the operation and management of power systems. Since power load is affected by a variety of complex factors (such as past net load values, time, weather, etc.), and these factors are in dynamic change, load forecasting becomes a challenging task. Traditional forecasting methods are difficult to effectively deal with the complexity and uncertainty of power load and the nonlinear relationship between multiple influencing factors. Neural networks have strong learning and generalization capabilities, can handle nonlinear relationships and adapt to dynamic changes, and are a powerful tool for solving short-term net load forecasting problems. However, there is currently no perfect short-term net load forecasting method based on neural networks that covers comprehensive technical links (from data preprocessing to model optimization). Summary of the invention

[0003] In order to solve the technical problem that there is no perfect short-term net load forecasting method based on neural network and covering comprehensive technical links (from data preprocessing to model optimization) in the above background technology, the present invention provides a short-term net load forecasting method;

[0004] Including: S1, data preprocessing: data preprocessing based on historical net load data, decomposed net load data, weather data, date data and other factors, the data preprocessing technology includes data cleaning, feature selection and feature scaling, wherein: data cleaning is used to process errors, missing or invalid data in the data set, including deleting duplicate data, processing missing data and outliers;

[0005] S2. Feature selection: Select features that are meaningful to the problem from the original data through filtering, wrapping, and embedding methods;

[0006] S3, Feature Scaling: Convert features of different scales to similar scales, including standardization and normalization methods;

[0007] S4. Construction of a net load forecasting model based on a neural network: including studying the characteristics and basic principles of short-term net load forecasting, including inaccuracy, conditionality and timeliness;

[0008] S5. Design the neural network structure according to the prediction characteristics: select factors reflecting load changes as input, net load prediction value as output, determine the training sample set and output set, and analyze the input parameters and hierarchical structure of the model;

[0009] S6. Build a net load prediction model and calculate the prediction value through the forward propagation algorithm of the neural network;

[0010] S7. Training and evaluation optimization of the net load prediction model: including: using appropriate training methods to tune and train the prediction model, calculating the error between the output value and the actual output value, modifying the connection weights of the network according to the error, and adjusting the learning rate and the number of training times;

[0011] S8. Select appropriate evaluation indicators to evaluate the prediction model: the evaluation indicators include mean absolute error and root mean square error. The model performance is evaluated by preparing test data, calculating evaluation indicators, and analyzing results.

[0012] The model is optimized based on the evaluation results, including adjusting the network structure, retraining and testing the model until satisfactory performance is achieved.

[0013] Data preprocessing based on multiple factors

[0014] Data cleaning is a key step in the data preparation stage. For possible duplicate data, use the deletion operation to ensure the uniqueness of each data point in the data set. For missing data, choose the processing method according to the specific situation. If the missing data sample has little impact on the overall data distribution, directly delete the sample with missing data; if the impact is large, use a filling method such as mean filling, median filling or model-based filling (such as using a regression model to predict missing values). For outliers, use statistical methods (such as the 3-sigma principle) or visualization methods (such as box plots) to identify them, and then handle them according to the actual situation, such as directly deleting them or using reasonable interpolation methods to correct them.

[0015] Feature Selection This paper studies various feature selection methods to improve model accuracy and reduce computational cost.

[0016] Filtering method: Calculate the correlation coefficient (such as Pearson correlation coefficient) between each feature in the original data and the target variable (net load value), sort the features according to the correlation, and select the features with higher correlation. For example, for net load prediction, weather temperature may have a high correlation with net load, while some features with weaker correlation with net load can be excluded.

[0017] Wrapping method: Using the performance of the prediction model as the evaluation criterion, gradually select feature subsets from the original feature set. For example, first randomly select a feature subset, use the subset to train the neural network model, and then gradually increase or decrease features based on the performance of the model on the validation set (such as mean square error), retrain the model and evaluate the performance until the optimal feature subset is found.

[0018] Embedding method: During the training of the neural network model, regularization methods (such as L1 or L2 regularization) are used to select features. The regularization term will penalize the weights of the model, making the weights corresponding to unimportant features approach 0, thereby achieving feature selection;

[0019] Feature Scaling

[0020] Standardization: Let the mean of feature x be μ and the standard deviation be σ, then the standardized feature This method scales the features to a normal distribution with a mean of 0 and a standard deviation of 1, which helps improve the training efficiency and stability of the model.

[0021] Normalization: For feature x, if its minimum value is χmin and its maximum value is χmax, then the normalized feature

[0022] Normalization scales the features to between 0 and 1, preserving the original distribution of the features, which is beneficial to model training in some cases.

[0023] Construction of Net Load Forecasting Model Based on Neural Network

[0024] Research on prediction characteristics and principles In-depth study of the characteristics of short-term net load forecasting, including forecast inaccuracy (affected by a variety of complex factors), conditionality (depending on time, weather and other conditions) and timeliness (short-term forecast in hours). According to these characteristics, the basic principles of forecasting are determined based on the inherent relationship between historical net load data, weather data, date data and other factors. For example, the change law of net load in different time periods (such as working days and holidays) and different weather conditions (such as sunny and rainy days) is analyzed to provide a theoretical basis for model construction.

[0025] Neural network structure design According to the characteristics of load forecasting, the neural network structure is designed. The factors that best reflect load changes (such as pre-processed historical net load data, temperature and humidity in weather data, whether it is a holiday in date data, etc.) are selected as the input of the neural network, and the net load forecast value is used as the output. Determine the appropriate training sample set and output set, and analyze the input parameters and hierarchical structure of the neural network model. For example, for a multi-layer perceptron (MLP) neural network, the number of input layer neurons can be determined according to the number of input features, the number of hidden layers and the number of neurons in each layer can be determined according to the complexity of the forecasting task, and the output layer is a single net load forecast value.

[0026] Net load forecasting model construction

[0027] Based on the above analysis, a net load prediction model is constructed. The historical net load data, weather data, date data and other pre-processed multi-factor data are used as input, and the prediction value is calculated by the forward propagation algorithm of the neural network. Assuming that the input vector of the neural network is χ=(χ1,χ2,…,χn), the weight matrix is ​​W, the bias vector is b, and the activation function is f(·), the calculation in the hidden layer is h=f(W1x+b1), and the calculation in the output layer is in is the predicted net load value.

[0028] Optimization techniques for training and evaluating net load forecasting models

[0029] Model tuning training technology The present invention selects a suitable training method (such as the back propagation algorithm in supervised learning) according to the specific situation. During the training process, the input sample set is input into the neural network, and the error between the output value and the actual output value is calculated.

[0030] Assume that the actual output is y and the predicted output is , then the error function can be the mean square error (MSE), that is Where m is the number of samples. According to the error, the connection weights of the network are modified according to certain rules (such as gradient descent method). For the gradient descent method, the weight update formula is Where a is the learning rate. Due to the complex structure of the neural network and the large number of neurons, it is necessary to appropriately increase the number of training times and adjust the learning rate to ensure that the model converges to better performance.

[0031] Model Evaluation Techniques

[0032] Test data preparation: Select a test data set that is independent of the training data and representative. The data set contains decomposed historical net load data and combines weather changes, holidays and other data to ensure that it can fully reflect the performance of the model in various situations.

[0033] Evaluation indicator selection: According to the characteristics of power load forecasting, select appropriate evaluation indicators, such as mean absolute error (MAE), Root mean square error (RMSE), And accuracy (if there is a classification task related situation), etc.

[0034] Testing and indicator calculation: Use the test data set to test the trained neural network model, record the prediction results of each test sample, and then calculate the selected evaluation indicators based on the prediction results and the true labels.

[0035] Result analysis and interpretation: Compare the results of different evaluation indicators to fully understand the performance of the model. Analyze which categories or features the model performs better or worse, and find out possible reasons and improvement directions.

[0036] Model Optimization

[0037] Optimize and adjust the model based on the evaluation results. For example, if it is found that the model has a large prediction error in certain time periods or weather conditions, the network structure can be adjusted, such as increasing the number of neurons in the hidden layer, changing the activation function, or adjusting the number of layers. Then retrain and test the model, and repeat this process until the model achieves satisfactory performance.

[0038] The beneficial effects of the present invention are: 1. Improving prediction accuracy

[0039] The positive impact of data preprocessing Through data cleaning, errors, missing and invalid data in the data can be effectively removed. For example, by processing outliers, the interference of abnormal data on the prediction model is avoided. In actual power load data, outliers may be caused by equipment failure, data collection errors, etc. Removing these outliers can enable the model to learn the normal load change law more accurately.

[0040] Feature selection methods can select the most relevant features for net load forecasting from a large number of raw data features. For example, by using the wrapping method to select features based on model performance, features that have little impact on the forecast results or even interfere with them can be excluded, allowing the model to focus on key factors and improve forecast accuracy.

[0041] Feature scaling converts features of different scales to similar scales, which helps the neural network model converge faster and improves prediction accuracy. The scale differences of different features may cause the model to over-focus on some features and ignore other features during training. Feature scaling can avoid this and enable the model to utilize all feature information more evenly.

[0042] Adaptability of Neural Network Model Construction This paper deeply studies the characteristics of short-term net load forecasting (inaccuracy, conditionality and timeliness) to construct a neural network model, so that the model structure is highly adapted to the net load forecasting task. For example, by selecting factors closely related to net load changes (such as weather data, date data, etc.) as input, it is possible to accurately capture the changing patterns of net load under different conditions, thereby improving the accuracy of the forecast.

[0043] The hierarchical structure and parameters of the neural network are determined according to the complexity of the prediction task, avoiding the model being too simple or too complex. The appropriate network structure can better fit the nonlinear relationship in the net load data, thereby more accurately predicting the future net load value.

[0044] Guarantee of model training and evaluation optimization During the model training process, selecting appropriate training methods and adjusting network connection weights according to errors can continuously optimize the model and improve its ability to fit the net load data. For example, by using the back propagation algorithm in supervised learning, the model can gradually learn accurate prediction rules by continuously reducing the error between the predicted output and the actual output (such as mean square error).

[0045] Selecting multiple evaluation indicators (such as mean absolute error, root mean square error, etc.) to comprehensively evaluate the model can accurately measure the performance of the model. These evaluation indicators reflect the difference between the model prediction results and the true value from different angles. Optimizing the model based on the evaluation results can further improve the prediction accuracy of the model.

[0046] 2. Improve model generalization ability

[0047] Data preprocessing improves generalization ability Feature selection and scaling operations in data preprocessing help improve the generalization ability of the model. By selecting representative features, the model can better adapt to different data sets, while feature scaling avoids overfitting of the model to features of a specific scale, making the model more stable in predicting new data.

[0048] The method of handling missing data and outliers during data cleaning ensures the quality and integrity of the model training data, which helps the model to effectively learn on a wider range of data and thus improve generalization capabilities.

[0049] The influence of neural network structure and training. A reasonable neural network structure is constructed, and the input, output and hierarchical structure are determined according to the prediction task, so that the model can have good generalization ability under different load scenarios. For example, in different seasons, weather conditions and date types, the model can accurately predict the net load based on the learned rules.

[0050] During the model training process, adjusting parameters such as learning rate and number of trainings, as well as using techniques to prevent overfitting (such as early stopping) can help improve the generalization ability of the model. Avoiding overfitting can enable the model to perform well on both training and test data, thereby effectively predicting different power load conditions in practical applications.

[0051] 3. Enhance the adaptability and flexibility of the model

[0052] Adaptability of data preprocessing The data preprocessing technology of the present invention has strong adaptability. In data cleaning, the processing methods for missing data and outliers can be selected and adjusted according to the actual situation of the data. For example, for power load data of different sources or qualities, the most suitable missing data processing method (delete samples or fill) and outlier processing method (delete or correct) can be flexibly selected.

[0053] The diversity of feature selection methods (filtering, wrapping, embedding) enables the model to adapt to different data characteristics and prediction needs. Different data sets may have different feature importance distributions. These methods can select the most effective feature combination according to the specific situation, thereby improving the adaptability of the model to different data.

[0054] Feature scaling methods (standardization and normalization) can also be selected based on the distribution characteristics of the features. This flexibility ensures that the model can be effectively preprocessed under different data distributions.

[0055] Flexibility in model construction and optimization When building a net load forecasting model, the structure and activation function of the neural network can be adjusted according to the actual forecasting effect. For example, different activation functions can be selected according to the nonlinearity and complexity of the data, or the number of hidden layers and neurons of the neural network can be increased or decreased according to the requirements of the forecasting task. This flexibility enables the model to adapt to different power load forecasting scenarios.

[0056] In the process of model training and evaluation optimization, there are many options for training methods, evaluation indicators, and optimization strategies. You can choose the appropriate training method based on factors such as data scale and hardware resources, select evaluation indicators based on different evaluation requirements, and adopt different optimization strategies based on the performance of the model, which enables the model to be effectively built, trained, and optimized in various practical application scenarios.

[0057] 4. Improve data utilization efficiency

[0058] Data utilization optimization and data cleaning in data preprocessing ensure the quality of data, so that valid data can be fully utilized. After removing erroneous, missing and invalid data, the remaining data can better reflect the actual situation of power load, thereby improving the utilization value of data.

[0059] Feature selection reduces the interference of irrelevant features on model training and improves data utilization efficiency by screening out features related to net load prediction. The model can focus on learning the relationship between key features and net load, avoiding wasting computing resources on useless information.

[0060] Data utilization in model building and training

[0061] When building a neural network model, selecting the factors that best reflect load changes as input can fully utilize these key data for accurate prediction. At the same time, reasonably determining the training sample set and output set, and using appropriate training methods to effectively learn the sample data, improves the efficiency of data utilization in the model building and training process.

[0062] 5. Contribute to the efficient operation and management of the power system

[0063] Accurate forecasting supports the operation of power systems. Accurate short-term net load forecasting can provide an important basis for the formulation of power generation plans for power systems. Power companies can reasonably arrange the start and stop of power generation equipment and the power generation capacity based on the forecast results, avoid insufficient or excessive power generation, improve power generation efficiency, and reduce operating costs.

[0064] For the power dispatching department, accurate net load forecasting helps optimize power grid dispatching, adjust the power grid's power flow distribution in advance, ensure the stability and reliability of power supply, and reduce the occurrence of power outages.

[0065] Optimizing resource allocation and cost control Through accurate net load forecasting, the power system can better allocate resources. For example, it can reasonably arrange the charging and discharging of energy storage equipment, improve the utilization efficiency of energy storage equipment, and also help integrate distributed energy resources and better integrate them into the power system.

[0066] In the power market environment, accurate net load forecasting can help power suppliers and users formulate reasonable trading strategies, reduce market risks, control costs, and promote the healthy development of the power market. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 : Module diagram of the present invention. DETAILED DESCRIPTION

[0068] Short-term net load forecasting is to infer the load value at a certain moment in the future based on the past net load value of the power load and time, weather, etc., so the object of load forecasting work is uncertain events. It is affected by a variety of complex factors, and various influencing factors are also developing and changing. Various load forecasts are made under certain conditions. As for the conditions, they can be divided into necessary conditions and assumed conditions; various load forecasts have a certain time range. Because load forecasting belongs to the category of scientific forecasting, it requires a relatively accurate quantitative concept and often needs to specify the forecast time. Therefore, it is necessary to select a suitable forecasting model.

[0069] Neural networks have powerful learning and generalization capabilities. They are able to learn complex nonlinear relationships from a large amount of historical data and make accurate predictions based on them. This ability enables neural networks to handle the complexity and uncertainty in power load forecasting. Neural networks can flexibly introduce multiple influencing factors, such as weather, seasons, holidays, etc. These factors have a significant impact on power load. Taking these factors into account through neural networks can improve the robustness and accuracy of the forecasting model. Neural network models can update parameters in real time and continuously adjust the forecast results based on the latest data. This real-time performance enables neural networks to adapt to dynamic changes in the power system and meet the real-time requirements in practical applications. Therefore, this project chooses neural network technology to achieve short-term net load forecasting.

[0070] First, the data preprocessing technology based on decomposed historical net load data, weather data, date data and other factors is studied. The data is abstracted and simplified by means of data cleaning, feature selection, data scaling and other means to ensure the accuracy and reliability of the data, reduce random fluctuations in the data, and improve model training efficiency and prediction accuracy. Secondly, the model technology based on the net load prediction neural network is studied. By combining the characteristics of net load prediction with the neural network model, the neural network structure is determined based on a large number of non-structural and non-precise rules of net load prediction. In the case of insufficient load data samples and features, the topological structure parameters and layer structure are designed to achieve short-term net load prediction with hours as the minimum unit. Finally, the net load prediction model training and evaluation technology is studied, and the hyperparameter tuning training is performed using the decomposed data, and the model evaluation indicators are calculated, and the model is tuned according to the evaluation results.

[0071] Research on data preprocessing technology based on historical net load data, decomposed net load data, weather data, date data and other factors

[0072] Due to the complexity and uncertainty of power load data, modeling and prediction using only raw data often cannot achieve satisfactory results. Data preprocessing methods can process and optimize the original load data through data cleaning, feature extraction, data scaling and other means, thereby improving the accuracy and reliability of load forecasting. Data preprocessing includes three main steps. First, the data cleaning process is studied. Load data often contains noise and outliers, which have a great impact on the accuracy of load forecasting. Data cleaning methods can improve the quality of load data by removing outliers and repairing missing values. Secondly, the data feature selection method is studied. By studying filtering, wrapping, embedding and other methods, reasonable features are selected and extracted to effectively reflect the characteristics and change laws of load data. Finally, the data feature scaling method is studied to convert features of different scales into similar scales. The quality of data preprocessing directly affects the performance and accuracy of subsequent models. Data preprocessing is an iterative process. This project needs to continuously try and adjust the preprocessing steps to obtain the best data processing effect.

[0073] Construction of Net Load Forecasting Model Based on Neural Network

[0074] Neural network is a very promising forecasting method. Its main advantage is that it can learn a large amount of data and approximate any nonlinear function. To this end, this project studies the forecasting model for short-term load forecasting. First, the research is conducted on the characteristics of forecast inaccuracy, conditionality, and timeliness, and then the basic principles of forecasting are studied to provide a basis for the construction of the model. Secondly, the data-driven neural network structure is studied in combination with the characteristics and principles of load forecasting, and the appropriate parameters and hierarchical structure are evaluated to improve the adaptability of the neural network structure to the load forecasting scenario. Finally, the construction technology based on the neural network structure is studied in combination with the characteristics of load forecasting to form a short-term net load forecasting model to predict the load value at similar moments in the future.

[0075] Optimizing the training and evaluation of the net load forecasting model

[0076] The training and evaluation of the model is a process of performance optimization and capability verification. First, study effective training methods. There are many training methods for machine learning models, including supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Different methods are suitable for different types of tasks and data. In this topic, select appropriate training methods according to specific problem settings and data conditions to improve training efficiency. Secondly, study the evaluation technology of the prediction model, select appropriate evaluation indicators for the type of neural network and specific load forecasting scenarios and needs, and support multiple aspects such as model performance measurement, tuning, risk control, and continuous improvement. Finally, study model optimization technology, improve training speed and model accuracy by adjusting the learning rate and gradient update strategy, accelerate convergence and prevent gradient disappearance or explosion by adjusting the weight distribution, and improve model performance through hyperparameter optimization.

[0077] Embodiment 1, data preprocessing implements data collection: collects historical net load data, weather data (such as temperature, humidity, weather type, etc.), date data (such as whether it is a working day, whether it is a holiday, etc.) and other data.

[0078] Data cleaning: Clean the collected data. Use database query statements or data processing software to identify and delete duplicate data points. For missing data, if the missing proportion of net load data in a certain time period is small, use the mean of the net load in that time period to fill in the missing values; for missing values ​​in weather data, use linear interpolation to fill in according to the time series characteristics of weather data. Identify outliers in net load data by drawing box plots. For outliers that obviously deviate from the normal range, if they are isolated points, delete them directly. If they are continuous abnormal intervals, make corrections based on the trend of surrounding normal data.

[0079] Feature selection: First, the filtering method is used to calculate the Pearson correlation coefficient between each feature and the net load value, and the features with an absolute correlation value greater than 0.3 are selected as the preliminary screening results. Then, the wrapping method is used, and the mean square error is used as the evaluation criterion. Based on the preliminary screening results, features are gradually added or deleted from the feature subset. After each addition or deletion of a feature, a simple neural network model is retrained and the mean square error is evaluated. After multiple iterations, the optimal feature subset is determined. Finally, the embedding method is used to further screen the features through L2 regularization during the complete neural network model training process to ensure the simplicity and effectiveness of the model.

[0080] Feature scaling: Select the feature scaling method based on the distribution characteristics of the data. For net load data and temperature data, since their distribution is close to normal distribution, the standardization method is used for scaling; for date data (after encoding) and humidity data, the normalization method is used for scaling.

[0081] Forecast model construction and implementation forecast characteristics and principle analysis: Analyzing the historical net load data, it is found that there are obvious differences in the net load curves on weekdays and holidays, and weather changes also have a significant impact on the net load. Based on these characteristics, the forecast principle based on time (weekdays / holidays), weather temperature, humidity and other factors is determined, that is, under different time and weather conditions, the net load will follow certain regular changes.

[0082] Neural network structure determination: A three-layer multi-layer perceptron (MLP) neural network is constructed as a net load prediction model. The input layer contains 6 neurons after feature selection, corresponding to time coding, temperature, humidity, net load in the previous hour, net load at the same time of the previous day, and season coding. The hidden layer is set to one layer, containing 12 neurons, and the activation function uses the ReLU function. The output layer is a single neuron, representing the net load prediction value, and the activation function uses a linear function.

[0083] Model construction: According to the above neural network structure, initialize the weight matrix and bias vector. The weight matrix is ​​randomly initialized, and the bias vector is initialized to 0. Then write code to implement the forward propagation algorithm of the neural network, calculate according to the formula mentioned above, and build a net load prediction model.

[0084] Model training and evaluation optimization implementation Model tuning training: Use the back propagation algorithm in supervised learning to train the constructed neural network model. Divide the training data set into a training set and a validation set with a ratio of 8:2. Set the initial learning rate to 0.01, calculate the mean square error for each iteration, and adjust the learning rate according to the change of the mean square error. If the mean square error does not decrease for 5 consecutive iterations, reduce the learning rate to 0.5 times the original. Update the weights according to the weight update formula, and set the number of training times to 1000. During the training process, evaluate the model performance on the validation set every 100 iterations and record the change of the mean square error.

[0085] Model evaluation: Prepare a test data set that contains net load data in different seasons and weather conditions. Select mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators. Use the test data set to test the trained model and calculate the values ​​of MAE and RMSE according to the calculation formula of the evaluation indicators. For example, after testing, MAE = 5.2 and RMSE = 7.1.

[0086] Model optimization: According to the evaluation results, if the RMSE value is large, it means that the model's prediction results have large fluctuations. Optimize the model, increase the number of hidden layer neurons to 16, reinitialize the weights and biases, and then train and test again. Repeat this process until the MAE and RMSE values ​​meet the pre-set requirements (such as MAE < 3, RMSE < 5).

Claims

1. A short-term net load forecasting method based on neural network, characterized in that: include: S1. Data preprocessing: Data preprocessing based on historical net load data, decomposed net load data, weather data, date data and other factors. The data preprocessing technology includes data cleaning, feature selection and feature scaling, wherein: data cleaning is used to process errors, missing or invalid data in the data set, including deleting duplicate data, processing missing data and outliers; S2. Feature selection: Select features that are meaningful to the problem from the original data through filtering, wrapping, and embedding methods; S3, Feature Scaling: Convert features of different scales to similar scales, including standardization and normalization methods; S4. Construction of a net load forecasting model based on a neural network: including studying the characteristics and basic principles of short-term net load forecasting, including inaccuracy, conditionality and timeliness; S5. Design the neural network structure according to the prediction characteristics: select factors reflecting load changes as input, net load prediction value as output, determine the training sample set and output set, and analyze the input parameters and hierarchical structure of the model; S6. Build a net load prediction model and calculate the prediction value through the forward propagation algorithm of the neural network; S7. Training and evaluation optimization of the net load prediction model: including: using appropriate training methods to tune and train the prediction model, calculating the error between the output value and the actual output value, modifying the connection weights of the network according to the error, and adjusting the learning rate and the number of training times; S8. Select appropriate evaluation indicators to evaluate the prediction model: the evaluation indicators include mean absolute error and root mean square error. The model performance is evaluated by preparing test data, calculating evaluation indicators, and analyzing results. The model is optimized based on the evaluation results, including adjusting the network structure, retraining and testing the model until satisfactory performance is achieved.

2. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: During data preprocessing, the order of data cleaning, feature selection, and feature scaling can be adjusted according to the characteristics of the data and actual needs, and each step can be iterated multiple times to achieve the best data preprocessing effect, thereby improving the accuracy and generalization ability of the model.

3. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: When constructing the net load prediction model, the activation function of the neural network is not limited to the ReLU function and the linear function. Other suitable activation functions, such as the sigmoid function and the tanh function, can be selected according to the actual prediction effect. Different activation functions can be used for different hidden layers to adapt to different data characteristics and prediction tasks.

4. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: In the model training and evaluation optimization technology, in addition to the back-propagation algorithm, other effective training algorithms can be selected according to the scale, characteristics and hardware resource conditions of the data, such as stochastic gradient descent (SGD) and its variants (Adagrad, Adadelta, Adam), and early stopping can be used during the training process to prevent overfitting and improve the generalization ability of the model.

5. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: When optimizing the model based on the evaluation results, in addition to adjusting the network structure, retraining and testing the model, the data preprocessing process can also be adjusted, such as reselecting features, changing feature scaling methods, and based on the model's performance in different time periods and weather conditions, partial structure or parameters of the model can be optimized in a targeted manner to improve the model's prediction performance in specific scenarios.

6. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: When selecting factors that reflect load changes as inputs to the neural network, the selection of these factors should be adjusted according to the characteristics of different regions and different power systems, and the input factors can be dynamically increased or decreased based on the actual data mining results to ensure that the model can accurately reflect the changing patterns of local power load.

7. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: When determining the training sample set and the output set, the division ratio of the training sample set (such as the ratio of the training set and the validation set) can be flexibly adjusted according to the quantity and quality of the data and the complexity of the model, and cross-validation technology can be used to improve the accuracy and reliability of model evaluation.

8. The short-term net load forecasting method based on neural network according to claim 1 is characterized in that: During the data preprocessing process, the model-based filling method for missing data is not limited to using regression models to predict missing values. Other machine learning models or statistical methods can also be used for filling. When dealing with outliers, the method for correcting outliers can be customized according to the distribution characteristics of the data and the requirements of the prediction task.

9. The short-term net load forecasting method based on neural network according to claim 1, characterized in that: In model evaluation techniques, in addition to mean absolute error (MAE) and root mean square error (RMSE), other evaluation metrics can be selected, such as mean absolute percentage error (MAPE), which is calculated as Or the Symmetric Mean Absolute Percent Error (SMAPE), and a variety of evaluation indicators can be combined to comprehensively evaluate the model performance according to different evaluation requirements; The meaning of each part is as follows: n: sample size, representing the number of data points of predicted and actual values ​​involved in calculating the error; yi: actual value, in short-term net load forecasting, is the real net load data value, that is, the actual observed net load size at a specific moment; Prediction value: the net load prediction value at the corresponding time obtained by prediction methods such as neural network; The absolute difference between the actual value and the predicted value indicates the degree of deviation between the predicted value and the true value at a single data point, regardless of the direction, and only focusing on the numerical value; The relative error of a single data point reflects the deviation ratio of the predicted value relative to the actual value; The sum of the relative errors of all data points; Mean relative error, average the sum of relative errors of all data points to each data point; The mean absolute percentage error (MAPE) converts the average relative error into a percentage form in order to more intuitively measure the accuracy of the prediction method. The smaller the value, the smaller the deviation between the predicted value and the actual value, and the higher the accuracy of the prediction method.

10. The short-term net load forecasting method based on neural network according to claim 1, characterized in that: During the model training process, the strategy for adjusting the learning rate is not limited to the method of reducing the learning rate when the mean square error does not decrease for multiple consecutive iterations. Other strategies for dynamically adjusting the learning rate can also be formulated based on the model convergence speed and the trend of the prediction error to improve the efficiency and accuracy of model training. In data cleaning, when processing missing data, if the missing data samples have little impact on the overall data distribution, the samples containing missing data are directly deleted. If the impact is large, a filling method is used. The filling methods include mean filling, median filling or model-based filling. Model-based filling includes but is not limited to using linear regression models, decision tree regression models or neural network models to predict missing values. When using model-based filling, the data used to train the filling model should have similar distribution characteristics to the data set where the missing data is located to ensure the accuracy and rationality of the filling.

Images