Method for adaptively predicting prefecture and county loads based on data driving

Abstract

The invention discloses a prefecture and county load adaptive prediction method based on data driving. The method comprises the following steps: acquiring load data and meteorological data of a target area; screening strong correlation data according to the data features, dividing the data into a training set and a verification set, and establishing a patchtst prediction model for preliminary training; determining a hyper-parameter of the optimal prediction model through a particle swarm algorithm, determining dispersion of particle distribution and an average particle distance, and keeping an optimal position of a current particle swarm as an optimal parameter; performing prediction evaluation by taking the average absolute percentage error sum as a model evaluation index; an improved particle swarm algorithm is used to carry out adaptive optimization on the patchtst prediction model so as to improve the prediction precision and generalization ability of the model; through data driving and model optimization, adaptive prediction of load change is realized; the model can continuously learn and adjust parameters so as to adapt to the change of a load mode and maintain the accuracy of prediction.

Description

Technical Field

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

[0002] Accurate power load forecasting has an important impact on the production scheduling and planning of power systems. Power load forecasting has been a hot research direction since the reform and opening up, and how to build a power load forecasting model and improve the prediction accuracy of the power load forecasting model is the top priority in the field of power load forecasting. Nowadays, with the rapid development of artificial intelligence research, intelligent prediction models represented by machine learning algorithms and artificial neural networks are widely used in power load forecasting modeling problems. The powerful nonlinear modeling and learning capabilities of intelligent models are used to better deal with power load forecasting problems. In addition, for the prediction accuracy of power load forecasting models, one of the research focuses is to use intelligent optimization algorithms to optimize the relevant parameters of intelligent prediction models to improve prediction accuracy. In this regard, although a single power load forecasting model or a single power load forecasting model optimized by a standard optimization algorithm can achieve power load forecasting with a certain prediction accuracy, its prediction results often cannot meet the standards and requirements of modern prediction accuracy.

[0003] For example, a "power load forecasting method" disclosed in Chinese patent literature, with publication number: CN116191412A, discloses data acquisition; step two: determining the sequence periodic parameter M, performing periodic accumulation of the sampling sequence, and obtaining a new periodic accumulation sequence H1; step three: generating a load forecasting model; step four: estimating the model's pending parameter values; step five: obtaining a new forecasting model; step six: verifying the reliability of the load model. However, this solution still has the problem that the optimization results are limited by the periodic accuracy. Summary of the invention

[0004] In order to solve the problem of insufficient accuracy of existing power load forecasting, the present invention provides a data-driven county load adaptive forecasting method, which uses an improved particle swarm algorithm to adaptively tune the load forecasting model to improve the prediction accuracy and generalization ability.

[0005] In order to achieve the above object, the present invention provides the following technical solutions: A data-driven adaptive forecasting method for prefecture-level and county-level loads includes: obtaining load data and meteorological data of a target area; screening strongly correlated data according to data characteristics and dividing them into training sets and validation sets, and establishing a patchtst forecasting model for preliminary training; The hyperparameters of the optimal prediction model are determined by the particle swarm algorithm, the discreteness of the particle distribution and the average particle distance are determined, and the optimal position of the current particle swarm is retained as the optimal parameter; The prediction evaluation is performed through the mean absolute percentage error and as a model evaluation indicator. The improved particle swarm algorithm is used to adaptively tune the patchtst prediction model to improve the prediction accuracy and generalization ability of the model. It has the ability to process time series data, capture long-term trends and short-term fluctuations, which helps to predict load changes more finely. Through data-driven and model optimization, adaptive prediction of load changes is achieved. As new data is continuously added, the model can continue to learn and adjust parameters to adapt to changes in load patterns and maintain the accuracy of predictions. It can adapt to load forecasting tasks in different regions and different time scales.

[0006] Preferably, the hyperparameters include the number of neurons and the learning rate of the hidden layer in the prediction model. When the discreteness of the particle distribution and the average particle distance are both less than a threshold value, the particle swarm is judged to be in the optimal position and used as the optimal parameter. By monitoring the discreteness of the particle distribution and the average particle distance, it is possible to accurately judge whether the particle swarm has converged to the optimal position. When both indicators are less than the set threshold value, it means that the particle swarm has been closely gathered near the optimal solution, which helps to prevent the particle swarm from converging prematurely before reaching the true optimal solution, thereby ensuring the accuracy and reliability of the hyperparameter optimization.

[0007] Preferably, after acquiring the load data and meteorological data, data preprocessing is performed, and feature scaling is performed by data minimum and maximum normalization to scale different features of different data to the same range. This helps to eliminate the deviation caused by scale differences between data, so that the model can consider each feature more fairly, thereby improving the prediction or classification performance of the model. When the features are at similar scales, the step size of the gradient descent is more uniform, and the optimal solution can be found faster, thereby shortening the training time.

[0008] Preferably, when screening the strongly correlated data, the data features are evaluated by the Pearson correlation coefficient method to determine the correlation coefficient of different data based on covariance and standard deviation. The Pearson correlation coefficient eliminates the influence of data units, so that data of different dimensions can also be compared, and data features with strong correlation can be quickly located. By screening the strongly correlated data, redundant information in the data set can be reduced, and the quality and efficiency of the data can be improved.

[0009] Preferably, the patchst prediction model includes a forward process, in which a univariate sequence of length L is determined and the univariate sequence of length L is separated into M independent channel sequences. The local semantic information of the time series can be better preserved. Adjacent time points often have a high correlation, so treating them as a whole, i.e., a patch, helps the model capture this locality.

[0010] Preferably, the patchst prediction model includes dividing the independent channel sequence into a second sequence of length P, and generating a training sequence when any two consecutive second sequences have non-overlapping regions. It is possible to pay more attention to the local features in the time series, and by treating them as a whole, the model can better capture this locality, thereby taking into account the short-term dependencies of the time series during prediction. It is possible to flexibly process independent channel sequences of different lengths.

[0011] Preferably, in the patchtst prediction model, linear projection is used to generate a latent representation, the training sequence is mapped to a latent space of dimension D, and the temporal order is monitored by additive position coding. The model's ability to understand time series is enhanced. Through additive position coding, the model can assign a unique position vector to each time step, which is added to the latent representation of the sequence during training, thereby retaining the temporal order information. Since the latent representation is generated by linear projection, it is not limited by the length of the input sequence. This enables the PatchTST model to flexibly process time series data of different lengths without the need for additional adjustments to the model structure.

[0012] Preferably, in the patchst prediction model, a multi-head attention layer is provided to identify multiple vectors in the potential representation, generate a key matrix based on the key vector, generate a value matrix based on the vertical vector, and generate a query matrix based on the query vector, so that the model can integrate information from different perspectives, thereby generating a more comprehensive and accurate prediction result.

[0013] Preferably, in the patchst prediction model, a loss function based on mean square error is provided, and the loss based on M time series is determined and averaged, which helps to maintain the stability of the gradient during the training process.

[0014] Preferably, the patchst prediction model includes performing instance normalization, normalizing the time series generated based on the data to generate zero mean and unit standard deviation, so as to help alleviate the distribution shift effect between training and test data.

[0015] The present invention has the following advantages: (1) The improved particle swarm algorithm is used to adaptively tune the patchst prediction model to improve the prediction accuracy and generalization ability of the model. It has the ability to process time series data, capture long-term trends and short-term fluctuations, and helps to predict load changes more accurately. Through data-driven and model optimization, adaptive prediction of load changes is achieved. As new data is continuously added, the model can continue to learn and adjust parameters to adapt to changes in load patterns and maintain prediction accuracy. It can adapt to load prediction tasks in different regions and time scales; (2) It helps prevent the particle swarm from converging prematurely before reaching the true optimal solution, thereby ensuring the accuracy and reliability of hyperparameter optimization; (3) It enables the model to consider each feature more fairly, thereby improving the prediction or classification performance of the model. When the features are at a similar scale, the step size of the gradient descent is more uniform, and the optimal solution can be found faster, thereby shortening the training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described below are merely exemplary. A person skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0017] Figure 1 Schematic diagram of the prediction model logic in the embodiment.

[0018] Figure 2 1 is a flow chart of the prediction method in the embodiment.

[0019] Figure 3 3 is a schematic diagram comparing the prediction results of the chiller before and after optimization in the embodiment.

[0020] Figure 4 2 is a schematic diagram comparing prediction results of different models in the embodiment. DETAILED DESCRIPTION

[0021] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0022] In a preferred embodiment, the present invention discloses a data-driven adaptive forecasting method for county loads, including the following steps: obtaining load data and meteorological data of the target area; these data directly reflect the changes in regional power demand and external environment, providing a solid foundation for forecasting. Compared with traditional empirical models, data-driven methods can more accurately capture complex dynamic changes and improve the accuracy of forecasts.

[0023] According to the data features, the strongly correlated data is screened and divided into training set and validation set. Through data feature screening, irrelevant or weakly correlated data is eliminated, the noise of model training is reduced, and the generalization ability of the model is improved. At the same time, focusing on strongly correlated data can more directly reflect the key factors of load changes. The patchtst prediction model is established for preliminary training; it helps to predict load changes more precisely.

[0024] The particle swarm algorithm is used to determine the hyperparameters of the optimal prediction model. By simulating the foraging behavior of bird flocks, the optimal solution is found in the solution space, which can effectively avoid local optimality and improve the configuration efficiency of model parameters. This ensures that the prediction model runs under the optimal parameter configuration, further improving the prediction accuracy. The discreteness of particle distribution and the average particle distance are determined, solving the problem that the traditional particle swarm optimization algorithm is prone to premature closure. The optimal position of the current particle swarm is retained as the optimal parameter; adaptive tuning is achieved.

[0025] The forecast evaluation is performed through the mean absolute percentage error and as a model evaluation indicator. It can intuitively reflect the relative error between the predicted value and the actual value, and is one of the important indicators for evaluating the performance of the time series forecasting model. MAPE not only considers the absolute value of the error, but also the size of the error relative to the actual value. It is particularly suitable for application scenarios such as load forecasting that require high accuracy.

[0026] When using this solution, the improved particle swarm algorithm is used to adaptively tune the patchtst prediction model to improve the prediction accuracy and generalization ability of the model. It has the ability to process time series data, capture long-term trends and short-term fluctuations, and helps to predict load changes more precisely. Through data-driven and model optimization, adaptive prediction of load changes is achieved. As new data is continuously added, the model can continue to learn and adjust parameters to adapt to changes in load patterns and maintain prediction accuracy. It can adapt to load forecasting tasks in different regions and time scales.

[0027] In other embodiments, the hyperparameters described in this solution include the number of neurons and the learning rate of the hidden layer in the prediction model, and also include the Dropout layer in the prediction model to achieve adaptive tuning of key parameters in the prediction model. When the discreteness of the particle distribution and the average particle distance are both less than the threshold, by monitoring the discreteness of the particle distribution and the average particle distance, it is possible to accurately determine whether the particle swarm has converged to the optimal position. It is determined that the particle swarm is in the optimal position at this time and used as the optimal parameter. When both indicators are less than the set threshold, it means that the particle swarm has been closely gathered near the optimal solution, which helps prevent the particle swarm from converging prematurely before reaching the true optimal solution, thereby ensuring the accuracy and reliability of hyperparameter optimization.

[0028] In other embodiments, the present solution includes performing data preprocessing and performing feature scaling by data minimum and maximum normalization. It helps to eliminate the bias caused by scale differences between data, so that the model can consider each feature more fairly, thereby improving the prediction or classification performance of the model. Scale different features of different data to the same range. When the features are at a similar scale, the step size of the gradient descent is more uniform, and the optimal solution can be found faster, thereby shortening the training time. It helps to reduce the model's excessive reliance on individual features because all features are assigned the same weight range. This helps to improve the generalization ability of the model, reduce the risk of overfitting, and enable the model to perform better on unseen data.

[0029] When using load forecasting, it is necessary to comprehensively consider the influence of multiple factors. When fusing and analyzing data from different sources, feature scaling can eliminate inconsistencies between data and make the data fusion process smoother.

[0030] In other embodiments, the present solution includes evaluating data features by the Pearson correlation coefficient method, which eliminates the influence of data units, so that data of different dimensions can also be compared. Determine the correlation coefficient of different data based on covariance and standard deviation. Data features with strong correlation can be quickly located. By screening strongly correlated data, redundant information in the data set can be reduced, and the quality and efficiency of the data can be improved.

[0031] In other embodiments, the patchst prediction model includes a forward process, which significantly reduces the number of tokens input to the Transformer encoder. Including determining a univariate sequence of length L in the forward process, it is convenient to separate a sequence of any length into a fixed number of channel sequences. Separate the univariate sequence of length L into M independent channel sequences. The local semantic information of the time series can be better preserved. Adjacent time points often have a high correlation, so treating them as a whole, i.e., a patch, helps the model capture this locality.

[0032] In other embodiments, the patchtst prediction model in this solution includes dividing the independent channel sequence into a second sequence of length P, and the model can pay more attention to the local features in the time series. These local features are crucial for understanding the short-term fluctuations and trends of the time series, and help to improve the prediction accuracy of the model. When there are non-overlapping areas in any two consecutive second sequences, a training sequence is generated, and by treating them as a whole, the model can better capture this locality, thereby taking into account the short-term dependencies of the time series when predicting. Independent channel sequences of different lengths can be flexibly processed.

[0033] Each patch becomes an input token to the Transformer encoder instead of the entire time series. This greatly reduces the number of input tokens, thereby reducing the computational complexity and memory requirements of the model. This is especially important for processing long time series data, as traditional Transformer models are prone to computational resource and memory limitations when processing long sequences. Reducing the number of input tokens not only increases the training speed of the model, but also speeds up the inference of the model. This makes the PatchTST model more efficient in practical applications and can generate prediction results faster.

[0034] By adjusting the length P and step size S of the patch, the PatchTST model can flexibly process independent channel sequences of different lengths. This flexibility enables the model to be applied to a variety of practical scenarios to meet different needs. When it is necessary to process multivariate time series, the PatchTST model can be expanded simply by increasing the number of channels. Each channel can still be processed independently, and the sequence within each channel can be divided into multiple patches for processing. This expansion method does not significantly increase the computational complexity and memory requirements of the model.

[0035] Although each patch contains only local information, by concatenating multiple patches and inputting them into the Transformer encoder, the model can still capture long-term dependencies in the time series. This is because the Transformer's attention mechanism can automatically learn the connections between elements in the sequence and use these connections in the prediction task.

[0036] In other embodiments, the present solution includes using linear projection to generate a latent representation, mapping the training sequence to a latent space of dimension D, where higher dimensions can capture more detailed information, while lower dimensions help reduce computational complexity and the risk of overfitting, and enhance the model's ability to understand time series. Monitor temporal order through additive position coding. Through additive position coding, the model is able to assign a unique position vector to each time step, which is added to the latent representation of the sequence during training, thereby retaining temporal order information. Since the latent representation is generated by linear projection, it is not limited by the length of the input sequence. This enables the PatchTST model to flexibly process time series data of different lengths without additional adjustments to the model structure.

[0037] In other embodiments, in the patchst prediction model, a loss function based on mean square error is provided, which directly reflects the difference between the model prediction value and the true value. It includes determining the loss based on M time series and averaging them. It helps to maintain the stability of the gradient during the training process. It can achieve the simultaneous optimization of the prediction accuracy of M time series.

[0038] In another specific embodiment, the present solution discloses a data-driven adaptive forecasting method for prefecture-level and county-level loads, comprising the following steps: first, data collection is performed, wherein the original data of various parameters are obtained in the data collection, including collecting regional historical load data and weather data, wherein the weather data includes temperature data, humidity data, wind speed data, and irradiance data.

[0039] Then, data preprocessing is performed. The data preprocessing includes the following steps: first, data cleaning is performed, and the raw data is filtered through the data cleaning. This includes checking the data obtained from data collection to determine whether there are missing data, outliers, and duplicate data in the data set, and deleting the missing data, outliers, and duplicate data respectively.

[0040] After data preprocessing, data normalization is performed. Through the data normalization, the data ranges of different features are scaled to the same range, so as to avoid the data being unable to learn sequence information well due to large differences in dimensions, resulting in the model performance not being able to reach a relatively ideal level. This solution scales different features to the same range, usually between 0 and 1, through data minimum-maximum normalization (Min-MaxNormalization). The formula for the minimum-maximum normalization is as follows Wherein, Xnew is the normalized data, including load data and weather data, and the weather data includes temperature data, humidity data, wind speed data and irradiance data; X is the cleaned data, including load data and weather data, and the weather data includes temperature data, humidity data, wind speed data and irradiance data. Xmax and Xmin are the maximum and minimum values ​​of load data and weather data, respectively.

[0041] After the data is normalized, feature selection is performed. The factors affecting the prediction include temperature, humidity, wind speed, and irradiance. By determining the factors with strong correlation among all factors, it is helpful to improve the accuracy of the prediction, and it can reduce the complexity of the model and improve the calculation efficiency. This solution uses the Pearson correlation coefficient method to determine the factors with strong correlation affecting the prediction accuracy.

[0042] Among them, X is the normalized data including temperature, humidity, wind speed, irradiance, human perception, and different types of air conditioning loads; Y is the load data; cov(X,Y) is the covariance of X and Y; σ X is the standard deviation of X, σ Y is the standard deviation of Y.

[0043] After feature selection and determination of the influencing factors with strong correlation, the Patchtst prediction model is established. The Patchtst model structure is as follows Figure 1 As shown, it mainly includes 5 parts: forward process, patching, Transformer encoding stage, loss function, and instance normalization. When there is a set of multivariate time series data, the time series of future T values ​​(XL+1,...,XL+T) is predicted based on the lookback window L: (X1,...,XL), where each xt at time step T is a vector of dimension M.

[0044] like Figure 1 As shown, in the patchtst prediction model, after the univariate sequence is input, normalization and patching are performed, and then the projection and position embedding layer is entered. After the projection and position embedding, the Transformer encoder is entered. The Transformer encoder includes a multi-head attention module, and the multi-head attention module is connected to an addition or normalization layer, and the addition or normalization layer is connected to a feedforward neural network layer, and the feedforward neural network layer is connected to a second addition or normalization layer. The Transformer encoder is connected to a flattening layer and a fully connected layer. Finally, the univariate sequence is output.

[0045] The scheme also includes a forward process, which specifically includes the following steps: the i-th univariate sequence with a length of L starting from index 1 is represented as Where i = 1,…,M. Input (x1,…,x L ) is separated into M univariate sequences Each sequence enters the Transformer network according to the independent channel setting.

[0046] In other embodiments, the patching process includes the following steps: each input univariate time series x (i) First, it is divided into patches. These patches can be overlapping or non-overlapping. The patch length is recorded as P; the non-overlapping area between two consecutive patches is recorded as S. Then the patch process will produce a patch sequence Where N is the number of patches generated, which can be expressed as: Where P and S represent the patch length and the non-overlapping area between two consecutive patches, respectively.

[0047] In other embodiments, the Transformer encoder includes the following logic: PatchTST uses the original Transformer encoder, which maps the observed signal to the potential representation. Map the block to the Transformer latent space of dimension D while applying a learnable additive positional encoding To monitor the time order of blocks; In the formula Represents the input fed into the Transformer encoder.

[0048] Then, in multi-head attention, each head transforms the input into a query matrix, a key matrix, and a value matrix; In the formula, are query vector, key vector, and numerical vector respectively.

[0049] In addition, the attention output is obtained by proportional output Where softmax() is the activation function; dk is the dimension of vector k.

[0050] In other embodiments, the patchst prediction model includes a loss function. When used, in order to measure the error value between the prediction and the actual situation, MSE is used as the loss function. The following steps are included, in each of the above channels, the loss is averaged over M time series. The overall objective loss can be expressed as: In the formula, It is the predicted value in the period of L+1:L+T; is the true value in the period L+1:L+T; M is the time series.

[0051] In other embodiments, this solution includes the following steps when performing instance normalization. The number of neurons, drop-out and learning rate of the GRU hidden layer are optimized. The improved particle swarm optimization algorithm is used for solving. The improved particle swarm optimization algorithm introduces the average particle distance method to improve the probability of searching for the global optimum, and judges the current state of the population based on the overall change of particle fitness, which solves the problem that the traditional particle swarm optimization algorithm is easy to close prematurely. The calculation formula of the improved particle swarm optimization algorithm is as follows: Where D(t) is the discreteness of the particle distribution in the population; L is the maximum diagonal length of the search space; M is the number of particles; n is the dimension of the solution; x id is the d-dimensional coordinate of the i-th particle; is the average value of the d-dimensional coordinates of all particles; f i is the current fitness of particle i; is the average fitness of the population, f is the limit σ 2 The normalized scale factor of .

[0052] When σ 2 The smaller the particle size, the smaller the discreteness of the particles. When the number of iterations increases, D(t) and σ 2 Get smaller and smaller, when D(t) and σ 2 When it is less than a certain value, it is considered to enter the late search stage. Therefore, this patent retains the optimal position of the current particle swarm to guide the particles to jump out of the local optimum.

[0053] In other embodiments, the scheme includes establishing an evaluation index. The steps are as follows, using Mean Absolute Error (MAPE) and as model evaluation indicators, and the calculation process is as follows: Where n is the number of data; i represents the i-th data; Yi and yi are the actual value and predicted value of the load respectively.

[0054] In other embodiments, the present invention discloses a method for predicting regional air conditioning load regulation capacity, comprising the following steps: Figure 2 As shown in the figure, the input information of the model is first cleaned and normalized, and then used as the training set, validation set and test set of the proposed model. Secondly, patchst is used as the prediction model, and the improved particle swarm algorithm is used to optimize the selection of hyperparameters, so that the model can better adapt to the learning data and output high-quality loads.

[0055] When this scheme is used, various types of detailed data including historical load and weather data of a province from January 2022 to December 2023 are used. In order to obtain meaningful data features, the weather condition data most relevant to the load is selected and selected using the Pearson analysis method. These four features are selected as the main features of the input model as shown in the following table; variable temperature humidity Wind speed Irradiance Electrical load 0.72 0.54 0.38 0.31 Table 1.

[0056] Compare optimization algorithms, such as Figure 3As shown in the figure, in order to verify the effectiveness of the method proposed in this scheme, this scheme compares the optimized and unoptimized models. It can be seen that the unoptimized model has a low degree of fitting. After using the improved particle swarm optimization algorithm, the prediction model can well represent the actual load. The MAPE of the unoptimized model is 6.38%, and the MAPE of the optimized model is 4.57%, and the accuracy is improved by 28.36%.

[0057] To further verify the effectiveness of the prediction model in this paper, we compared the prediction curves of the two other models (GRU and transformer). Figure 4 As shown. Each prediction model is run with the same data set under the same hardware conditions, and the hyperparameters of the improved particle swarm selection model are adopted. Table 2 shows the prediction error of each model; Model MAPE(%) GRU 10.25 transformer 6.84 patchTST 4.52 Table 2.

[0058] Depend on Figure 4 From Table 2, we can see that the minimum MAPE value of the patchTST prediction model is 4.52%, and its prediction accuracy is the highest compared with the other two models. This shows that the prediction model proposed in this patent has higher accuracy and stronger generalization ability.

[0059] These modifications or improvements made on the basis of the present invention without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.

Claims

1. A data-driven adaptive forecasting method for county loads, characterized in that: include: Obtain load data and meteorological data for the target area; According to the data characteristics, the highly correlated data are selected and divided into training set and validation set, and the patchst prediction model is established for preliminary training; The hyperparameters of the optimal prediction model are determined by the particle swarm algorithm, the discreteness of the particle distribution and the average particle distance are determined, and the optimal position of the current particle swarm is retained as the optimal parameter; The predictions were evaluated by mean absolute percentage error and as model evaluation metrics.

2. According to claim 1, a data-driven adaptive forecasting method for county loads is characterized in that: The hyperparameters include the number of neurons in the hidden layer of the prediction model and the learning rate. When the discreteness of the particle distribution and the average particle distance are both less than the threshold, it is determined that the particle group is at the optimal position and is used as the optimal parameter.

3. A data-driven adaptive forecasting method for county loads according to claim 1 or 2, characterized in that: After the load data and meteorological data are obtained, data preprocessing is performed, and feature scaling is performed through minimum and maximum data normalization to scale different features of different data into the same range.

4. A data-driven adaptive forecasting method for county loads according to claim 1 or 2, characterized in that: The screening of strongly correlated data includes evaluating data characteristics through the Pearson correlation coefficient method to determine the correlation coefficient of different data based on covariance and standard deviation.

5. The data-driven adaptive forecasting method for county loads according to claim 1 is characterized in that: The patchst prediction model includes a forward process, in which a univariate sequence with a length of L is determined and the univariate sequence with a length of L is separated into M independent channel sequences.

6. A data-driven adaptive forecasting method for county loads according to claim 5, characterized in that: The patchst prediction model includes dividing the independent channel sequence into second sequences of length P, and generating a training sequence when any two consecutive second sequences have non-overlapping regions.

7. A data-driven adaptive forecasting method for county loads according to claim 1, 5 or 6, characterized in that: In the described patchtst prediction model, linear projection is used to generate latent representations, the training sequences are mapped into a latent space of dimension D, and the temporal order is monitored by additive positional encoding.

8. The data-driven adaptive forecasting method for county loads according to claim 7 is characterized in that: In the described patchst prediction model, a multi-head attention layer is provided to identify multiple vectors in the potential representation, generate a key matrix based on the key vector, generate a value matrix based on the vertical vector, and generate a query matrix based on the query vector.

9. A data-driven adaptive forecasting method for county loads according to claim 1, 5 or 6, characterized in that: In the patchst prediction model, a loss function based on mean square error is provided to determine the loss function based on M The losses on the time series are averaged.

10. A data-driven adaptive forecasting method for county loads according to claim 1 or 2 or 5 or 6, characterized in that: The patchtst prediction model includes instance normalization, which normalizes the time series generated based on the data to generate zero mean and unit standard deviation.

Images