Load prediction model dynamic adjustment method based on Hurst index guidance

Abstract

The invention discloses a load prediction model dynamic adjustment method based on Hurst index guidance, and relates to the technical field of power load prediction. According to the method, the long correlation of the data is judged by utilizing the Hurst index, so that whether the model is finely adjusted or not is guided, the model can be dynamically adjusted in real time according to the change of a data rule, the prediction precision and efficiency are improved, and a better engineering solution is provided for a load prediction task.

Description

Technical Field

[0001] The invention relates to the technical field of electric load forecasting, and in particular to a method for dynamically adjusting a load forecasting model based on Hurst exponent guidance. Background Art

[0002] Existing deep learning load forecasting methods usually use fixed parameter models, that is, after training, the structure and parameters of the model do not change. When the data distribution changes, the fixed parameter model is difficult to quickly adapt to the new data pattern, which affects the accuracy and stability of the forecast results. In addition, the application of the Hurst index in power systems is mostly limited to being used as a data feature statistic, which is only used to evaluate the stability or long-term correlation of time series, and has not yet been combined with the model dynamic adjustment mechanism to improve forecasting performance.

[0003] After searching the prior art, it was found that Chinese patent document No. CN119297995A, published on September 24, 2024, discloses a method for predicting power load, which integrates power load information, weather information and time characteristics, generates a feature coding sequence with load cycle change characteristics, weather characteristics, and fusion characteristics, and is used for predicting power load based on the Transformer framework. In addition, Chinese patent document No. CN118780449A, published on September 12, 2024, also proposed a method for predicting power load, including obtaining power system data, performing data fusion processing on the power system data, using the variational mode decomposition model to decompose the original time series into subsequences, constructing a power load prediction model according to the gated recurrent unit and the attention mechanism, and predicting the power load according to the decomposed subsequence using the power load prediction model, and using the improved sparrow search algorithm to optimize the model parameters, thereby improving the prediction accuracy. Compared with the present invention, the above-mentioned technology cannot solve the technical problem that the model cannot quickly adapt to the dynamic changes of data distribution, and cannot be dynamically adjusted according to real-time data, resulting in the reliability of the prediction results still needing to be improved.

[0004] Chinese patent document number CN110689169A, published on September 4, 2019, discloses a short-term power load forecasting method, in which the Hurst index is used to determine whether the sample data sequence has a long-term load correlation characteristic. Only when the Hurst index falls within the characteristic range will the system perform subsequent operations. This technology uses the Hurst index as an evaluation criterion for the quality of time series data, but is not further used to optimize the forecasting structure model. The application of the Hurst threshold is relatively simple, resulting in limited improvement in forecasting accuracy. Summary of the invention

[0005] The purpose of the present invention is to provide a method for dynamically adjusting a load prediction model guided by the Hurst exponent to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for dynamically adjusting a load prediction model guided by the Hurst exponent, comprising the following steps:

[0008] Step 1: Train a basic TCN-GRU prediction model through historical data, and the method is as follows:

[0009] 1.1) Preprocess the historical data: Structurally analyze the timestamp information, extract the time features of year, month, and day as independent dimensions, and standardize the load data, temperature data, and the parsed timestamp information data. Save the model for standardizing the load data as the load standardization model. At the same time, divide the standardized data into a training set and a test set according to a ratio of 0.828:0.172; Use the sliding window method to generate sample data, with an input length of 96 data points and an output of 16 future data points, and a sliding step size of 1 each time; 96 samples form a batch, thus processing the data into tensor data that can be processed by the deep learning model;

[0010] 1.2) Build a TCN-GRU model: The TCN-GRU model adopts a hybrid structure of TCN and GRU. The TCN branch adopts a stacked structure of dilated causal convolution, and captures multi-scale time series features through an exponentially expanding dilation coefficient; The GRU branch uses a gating mechanism to model the long-term dependence relationship of the sequence. Finally, the prediction result is output through a fully connected layer;

[0011] 1.3) Use the structured tensor data produced in 1.1) as the model input, use the mean square error MSE between the true value and the predicted value as the loss function, continuously optimize the model parameters during the training process, and select the optimal model on the test set for saving as the basic TCN-GRU prediction model;

[0012] Step 2: Calculate the Hurst exponent of the future load according to the real-time data, and select the rescaled range calculation method. The calculation process is as follows:

[0013] 2.1) Calculate using the load data of the latest 12 points, that is, the time series length N = 12. Take n = 2, 3,..., 12, and divide the sequence into m subintervals with a length of n. Each interval is denoted as where X k refers to the kth consecutive subinterval with a length of n, represents the nth load data in the kth consecutive subinterval; 2.2) For each subinterval X k:

[0014] First, perform mean detrending: where μ k is the mean of the k-th sub-interval, is the deviation of the i-th data point in the k-th sub-interval; then calculate the cumulative deviation sequence: where represents the cumulative deviation of the i-th point; Calculate the range: Calculate the standard deviation:

[0015] 2.3) Calculate the average rescaled range of all sub-intervals:

[0016] 2.4) Fit the power-law relationship by least squares: log(R / S) n = H·logn + C, where the slope H is the Hurst exponent;

[0017] Step 3: Judge the relationship between the calculated Hurst exponent and the threshold. If the Hurst exponent is greater than the threshold, directly proceed to Step 5; if the Hurst exponent is less than the threshold, first perform Step 4 and then Step 5;

[0018] Step 4: If in Step 3, the Hurst exponent is less than the threshold, trigger the model fine-tuning strategy, and the process is as follows:

[0019] 4.1) Based on the basic TCN-GRU prediction model trained in Step 1, fine-tune the last GRU layer and the fully connected layer in the model, and keep other weight parameters frozen;

[0020] 4.2) Taking the current time t as the benchmark, select the first 96 samples, set the training cycle Epoch to 5, and reduce the learning rate to 1e-4, so that the model can be quickly fine-tuned under limited computing resources. After training, select the optimal model as the new basic model to replace the basic TCN-GRU prediction model as the new basic TCN-GRU model;

[0021] Step 5: After completing the adaptive optimization of the model, use this model to predict the future load, and the specific implementation framework is as follows:

[0022] 5.1) After preprocessing the information of the past day, that is, 96 points, as the input of the new basic TCN-GRU model, obtain the output of the model, that is, the normalized predicted values of 16 points in the next 4 hours;

[0023] 5.2) According to the load normalization model in 1,1), denormalize the output data to obtain the true values predicted by the model.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] The present invention takes the Hurst exponent as the key criterion for judging whether the model needs to be fine-tuned. The Hurst exponent can effectively characterize the self-similarity and long-term dependence of time series. When the Hurst exponent is lower than the set threshold, the data shows strong anti-persistence or is close to a random walk, and the prediction accuracy of traditional models may be limited. Therefore, the present invention dynamically triggers the model fine-tuning mechanism based on the real-time calculated Hurst exponent, and uses the latest data to optimize the model parameters, so as to fully mine the potential information in the new data and improve the upper limit of the model prediction accuracy. Through the real-time monitoring and judgment of the Hurst exponent, the present invention can start the retraining process of the model in a timely and reasonable manner, ensuring that the model is always in the optimal state to cope with the complex and changeable environment in the power system.

[0026] The method proposed by the present invention combines the Hurst exponent and deep learning, and guides the adaptive adjustment of the model through the Hurst exponent, enabling the model to respond to data changes in real time. This method overcomes the limitation that traditional deep learning models are difficult to quickly adapt to changes in data distribution, making the model more flexible. Therefore, the present invention can effectively cope with the complexity and uncertainty of data in the power system, thereby improving the accuracy of model training and having important practical value in engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the flowchart of the dynamic adjustment of the load model guided by the Hurst exponent of the present invention;

[0028] Figure 2 It is the architecture diagram of the TCN-GRU neural network of the present invention;

[0029] Figure 3 It is the double logarithmic fitting diagram of the Hurst exponent calculation of the present invention;

[0030] Figure 4 It is the effect diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] Please refer to Figure 1 , in the embodiments of the present invention, a method for dynamically adjusting a load prediction model guided by the Hurst exponent includes the following steps:

[0033] Step 1: Train a basic TCN-GRU prediction model using historical data. The method is as follows:

[0034] 1.1) Preprocess the historical data: Structurally analyze the timestamp information, extract the year, month, and day time features as independent dimensions, and standardize the load data, temperature data, and the parsed timestamp information data. Save the model for standardizing the load data as the load standardization model. At the same time, split the standardized data into a training set and a test set at a ratio of 0.828:0.172. Use the sliding window method to generate sample data, with an input length of 96 data points and an output of 16 future data points, and a sliding step of 1 each time. 96 samples form a batch, thus processing the data into tensor data that can be processed by the deep learning model;

[0035] 1.2) Build a TCN-GRU model: The structure of the model is as Figure 2 shown. The TCN-GRU model adopts a hybrid structure of TCN and GRU. The TCN branch uses a stacked structure of dilated causal convolutions to capture multi-scale temporal features through an exponentially expanding dilation coefficient; the GRU branch uses a gating mechanism to model the long-term dependencies of the sequence. Finally, the prediction result is output through a fully connected layer;

[0036] 1.3) Use the structured tensor data made in 1.1) as the model input, use the mean squared error MSE between the true value and the predicted value as the loss function, continuously optimize the model parameters during the training process, and select the optimal model on the test set for saving as the basic TCN-GRU prediction model;

[0037] Step 2: Calculate the Hurst exponent of the future load based on real-time data. Select the rescaled range calculation method, and the calculation process is as follows:

[0038] 2.1) Calculate using the load data of the latest 12 points, that is, the time series length N = 12. Take n = 2, 3,..., 12, and divide the sequence into m consecutive subintervals of length n, and each interval is denoted as where X k refers to the kth consecutive subinterval of length n, represents the nth load data in the kth consecutive subinterval;

[0039] 2.2) For each subinterval X k :

[0040] First, detrend by the mean: where μ k is the mean of the kth subinterval, is the deviation of the i-th data point in the k-th sub-interval; then calculate the cumulative deviation sequence: where represents the cumulative deviation of the i-th point; range calculation: Standard deviation calculation:

[0041] 2.3) Calculate the average rescaled range of all sub-intervals:

[0042] 2.4) Fit the power-law relationship by least squares: log(R / S) n = H·logn + C, where the slope H is the Hurst exponent, as Figure 3 shown;

[0043] Step 3, judge the relationship between the calculated Hurst exponent and the threshold. If the Hurst exponent is greater than the threshold, directly go to Step 5; if the Hurst exponent is less than the threshold, first go to Step 4, and then go to Step 5;

[0044] Step 4, if in Step 3, the Hurst exponent is less than the threshold, trigger the model fine-tuning strategy, and the process is as follows:

[0045] 4.1) Based on the basic TCN-GRU prediction model trained in Step 1, fine-tune the last GRU layer and the fully connected layer in the model, and keep other weight parameters frozen;

[0046] 4.2) Taking the current time t as the benchmark, select the first 96 samples, set the training period Epoch to 5, and reduce the learning rate to 1e-4, so that the model can be quickly fine-tuned under limited computing resources. After training is completed, select the optimal model as the new basic model to replace the basic TCN-GRU prediction model as the new basic TCN-GRU model;

[0047] Step 5, after completing the adaptive optimization of the model, use this model to predict the future load, and the specific implementation framework is as follows:

[0048] 5.1) After preprocessing the information of the past day, that is, 96 points, as the input of the new basic TCN-GRU model, obtain the output of the model, that is, the normalized predicted values of 16 points in the next 4 hours;

[0049] 5.2) According to the load normalization model in 1,1), denormalize the output data to obtain the true values predicted by the model.

[0050] The load prediction method guided by the Hurst index is used to predict the load in multiple regions, and the MSE between the true value and the measured value has been significantly reduced. The MSE without model fine-tuning in a certain region is 0.0229. When fine-tuning the data with Hurst index less than 0.6, 0.7, 0.8, 0.9, and 1 respectively, the obtained MSEs are 0.0229, 0.0202, 0.0197, 0.0189, and 0.0075 respectively. The prediction results are as Figure 4 shown. It can be seen that the coincidence degree between the prediction curve and the true curve has been improved after fine-tuning. It can be seen that fine-tuning according to the Hurst index can effectively improve the accuracy of load prediction and has certain guiding significance.

[0051] Compared with the prior art and the static model (without fine-tuning), the prediction accuracy of this method has been significantly improved. Compared with complete online learning, this method saves computing resources, can perform real-time prediction, and has greater feasibility and efficiency. Technical verification shows that this method provides a better engineering solution for the time series prediction task in terms of accuracy, efficiency, and robustness.

[0052] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0053] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for dynamic adjustment of load forecasting model based on Hurst index guidance, characterized by: Train a basic TCN-GRU prediction model using historical data; Based on real-time data, the Hurst index of future load is calculated using the re-standard range calculation method; Determine the relationship between the calculated Hurst index and the threshold. If the Hurst index is greater than the threshold, the basic TCN-GRU prediction model is used directly for prediction. If the Hurst index is less than the threshold, the basic TCN-GRU prediction model is updated first, overwriting the basic TCN-GRU prediction model, and then prediction is performed.

2. The method for dynamic adjustment of load forecasting model based on Hurst index guidance according to claim 1, characterized in that: The process of training a basic TCN-GRU prediction model using historical data is as follows: Preprocess historical data into tensor data that can be processed by deep learning models; Build a TCN-GRU model; Use the processed tensor data to train the TCN-GRU model, and select the best model on the test set to save as the basic TCN-GRU prediction model.

3. The method for dynamic adjustment of load forecasting model based on Hurst index guidance according to claim 1, characterized in that: The historical data includes load data, temperature data and timestamp information data. The method for preprocessing the historical data is as follows: Perform structured analysis on the timestamp information, extract the time features of year, month and day as independent dimensions, standardize the load data, temperature data and the analyzed timestamp information data, and save the load data standardization model as the load standardization model; The standardized data is divided into training set and test set in the ratio of 0.828:0.172; The sliding window method is used to generate sample data, where the input data length is 96, the output data length is 16, the sliding step size is 1 each time, and 96 samples constitute a batch.

4. The method for dynamic adjustment of load forecasting model based on Hurst index guidance according to claim 1, characterized in that: The TCN-GRU model adopts a hybrid structure of TCN and GRU; the TCN branch adopts an expanded causal convolution stacking structure to capture multi-scale temporal features through an exponentially expanded expansion coefficient; the GRU branch uses a gating mechanism to model long-term sequence dependencies; finally, the prediction results are output through a fully connected layer.

5. The method for dynamic adjustment of load forecasting model based on Hurst index guidance according to claim 1, characterized in that: The method of calculating the Hurst index of future load using the re-standard range calculation method is as follows: Take the latest 12 points of load data for calculation, that is, the time series length N = 12, take n = 2, 3, ..., 12, divide the sequence into m continuous sub-intervals of length n, each interval is recorded as Where X k refers to the kth continuous subinterval of length n, represents the nth load data in the kth continuous subinterval; For each subinterval X k : First, detrend the mean: where μ k is the mean of the kth subinterval, is the deviation of the i-th data point in the k-th subinterval; then calculate the cumulative deviation sequence: in Indicates the cumulative deviation of the i-th point; range calculation: Standard deviation calculation: Calculate the average rescaled range of all subintervals: Fitting a power law relationship by least squares: log(R / S) n =H·logn+C, where the slope H is the Hurst exponent.

6. The method for dynamic adjustment of load forecasting model based on Hurst index guidance according to claim 1, characterized in that: The method for updating the basic TCN-GRU prediction model is as follows: Based on the basic TCN-GRU prediction model, if the Hurst index meets the conditions, the last GRU layer and the fully connected layer in the model are fine-tuned, and the other weight parameters remain frozen; Taking the current time t as the benchmark, the first 96 samples are selected, the training cycle Epoch is set to 5, and the learning rate is reduced to 1e-4, so that the model can be quickly fine-tuned under limited computing resources; After the training is completed, the optimal model is selected as the new basic model to replace the basic TCN-GRU prediction model as the new basic TCN-GRU model to update the basic TCN-GRU prediction model.

7. The method for dynamic adjustment of load forecasting model based on Hurst index guidance according to claim 1, characterized in that: The method of making predictions using the basic TCN-GRU prediction model is as follows: After preprocessing the information of 96 points in the past day, it is used as the input of the new basic TCN-GRU model to obtain the output of the model, that is, the normalized predicted value of 16 points in the next 4 hours; According to the load normalization model, the output data is denormalized to obtain the true value predicted by the model.

Images