Data-driven demand-side load forecasting and uncertainty estimation method and system

Through Bayesian deep learning and ProbSparse Transformer model, combined with the variational inference method, the problem of insufficient uncertainty estimation of traditional load prediction is solved, and the accuracy and reliability of load prediction are improved, providing more suitable energy management Support data of the letter.

CN118693818BActive Publication Date: 2024-12-06BEIJING INST OF TECH +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411162944.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-12-06
Estimated Expiration
2044-08-23

AI Technical Summary

Technical Problem

Traditional load prediction methods rely on statistics and time series analysis, with high data characteristics requirements and insufficient estimation of uncertainty. Deep learning methods have significant effects in load prediction, but still need to be improved for uncertainty estimation.

Method used

Introduce Bayesian deep learning, and model and train the Bayesian posterior distribution of neural networks through ProbSparse Transformer model and variational inference method to realize load prediction and uncertainty estimation, generate load prediction confidence intervals, and improve result reliability .

Benefits of technology

It improves the accuracy and credibility of load forecasting, generates efficient data support, provides more credible data for energy management, and achieves the goals of energy conservation, consumption reduction and environmental protection and emission reduction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118693818B_ABST
    Figure CN118693818B_ABST
Patent Text Reader

Abstract

The present application proposes a data-driven demand-side load forecasting and uncertainty estimation method and system, the method includes: collecting historical power load and corresponding weather data, and slicing the collected power load and weather data to construct a training sample set, on the basis of the ProbSparse Transformer model, constructing a Bayesian deep learning load forecasting model with distribution parameters, optimizing the model parameters using the variational inference method, and obtaining the target model after the training is completed; estimating the uncertainty and confidence interval of the load forecasting results through the target model; evaluating the prediction accuracy of the target model through indicator parameters, and verifying the uncertainty of the load forecasting results through the coverage judgment of the confidence interval. Based on the scheme proposed in this application, it can be applied to real-time online intelligent load forecasting, providing more accurate load forecasting and its uncertainty estimation, which is of great significance to improving the efficiency and accuracy of energy management, achieving energy saving and consumption reduction, and environmental protection and emission reduction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of load forecasting and energy system management, and in particular to a data-driven demand-side load forecasting and uncertainty estimation method and system. Background Art

[0002] Demand-side load forecasting and management play an important role in improving energy efficiency, enhancing energy dispatch and renewable energy absorption capacity, and also help reduce carbon emissions. Traditional methods mainly rely on statistical and time series analysis for load forecasting, but they have high requirements for data features and are limited in terms of confidence. With the development of deep learning technology, data-driven load forecasting methods have attracted much attention, especially methods based on models such as Transformer. Although they are effective in load forecasting, the estimation of uncertainty still needs to be improved. Summary of the invention

[0003] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0004] To this end, the first purpose of this application is to propose a data-driven demand-side load forecasting and uncertainty estimation method to introduce Bayesian deep learning into load forecasting, and to achieve intelligent and accurate prediction by modeling and training the Bayesian posterior distribution of neural networks. At the same time, this method can effectively estimate the uncertainty of the prediction results, generate load prediction confidence intervals, improve the reliability of the results, and provide more reliable data support for energy management. This method is expected to improve the efficiency and accuracy of energy management, achieve the goals of energy conservation and consumption reduction and environmental protection and emission reduction, and has broad application prospects in the fields of load forecasting and energy system management.

[0005] The second objective of this application is to propose a data-driven demand-side load forecasting and uncertainty estimation system.

[0006] To achieve the above objectives, the first embodiment of the present application proposes a data-driven demand-side load forecasting and uncertainty estimation method, including:

[0007] Collecting historical data, wherein the historical data includes historical power load data and historical weather data at corresponding times;

[0008] Slice the historical data into time series of fixed length using a sliding window, use the sliced ​​data as the input sequence, and use the load value after multiple steps in the future as the prediction label to construct a training sample set;

[0009] Based on the ProbSparse Transformer model, a Bayesian deep learning load forecasting model with distribution parameters is constructed, and the model parameters of the Bayesian deep learning load forecasting model are optimized by using a variational inference method according to the training sample set. After the training is completed, a target model is obtained;

[0010] Outputting a load forecast result according to the target model, and estimating the uncertainty and confidence interval of the load forecast result through the target model;

[0011] The prediction accuracy of the target model is evaluated through indicator parameters, and the uncertainty of the load prediction result is tested through coverage judgment of the confidence interval.

[0012] Optionally, the collecting of historical data, including historical power load data and historical weather data at corresponding moments, includes:

[0013] The sampling interval of the historical power load data shall not exceed 15 minutes;

[0014] The collection of the historical weather data takes into account the factor variables that affect the load energy consumption, and the factor variables include temperature, radiation, and humidity.

[0015] Optionally, the time range of the sliding window is not less than 3 hours.

[0016] Optionally, the sliced ​​data is used as the input sequence, and the load value after multiple steps in the future is used as the predicted label to construct a training sample set, including:

[0017]

[0018]

[0019]

[0020] in, represents the training sample set, is the total number of training samples, is the number of time steps contained in the sliding window, Reason The time series input vector consists of time steps. Respectively represent Samples at time , , The true value of Indicates power load data or corresponding historical weather data, To predict the label, Represents the prediction step size.

[0021] Optionally, based on the ProbSparse Transformer model, a Bayesian deep learning load forecasting model with distribution parameters is constructed, including:

[0022] Set the model parameters of the ProbSparse Transformer model, including the number of attention heads, the multi-head self-attention mechanism, and the number of feedforward network layers. The multi-head self-attention mechanism is used to extract features from the input sequence. The multi-head self-attention mechanism can adopt a sparse attention mechanism.

[0023] A model architecture of the ProbSparse Transformer model is set, wherein the model architecture includes an input layer, multiple Transformer encoding layers, multiple Transformer decoding layers, and an output layer. The output layer is used to output a load forecast result. The load forecast result includes a load forecast value and an uncertainty of the load forecast value, and the uncertainty is the variance and mean of the load forecast value.

[0024] Optional optimization of model parameters, including:

[0025]

[0026] in, is the variational distribution, is the true posterior distribution, represents the optimal parameters of the model, is the Kullback-Leibler divergence, which is used to measure the difference between the true posterior distribution and the variational distribution. is the weight parameter of the model.

[0027] Optionally, outputting a load forecast result according to the target model includes:

[0028] The other sliced ​​data are input into the target model as an input sequence to obtain a load forecast result.

[0029] Optionally, estimating the uncertainty and confidence interval of the load forecast result through the target model includes:

[0030] Through the variational distribution Expressing model uncertainty, the variational distribution Obtained through Monte Carlo sampling;

[0031] Express random uncertainty through Gaussian distribution of the output layer;

[0032] The confidence interval is constructed by the standard deviation and mean of the load forecast value, and the expression is:

[0033]

[0034] in, is the mean of the load forecast values, is the standard deviation of the load forecast value.

[0035] Optionally, the evaluating the prediction accuracy of the target model by using the indicator parameters and the checking the uncertainty of the load prediction result by using the coverage judgment of the confidence interval include:

[0036] Select a test sample set to perform model verification and evaluation on the prediction accuracy of the target model, and the indicator parameters include root mean square error, mean absolute error, and mean absolute percentage error;

[0037] The deviation between the load prediction value and the actual load value is calculated, and it is determined whether the confidence interval covers the actual load value. The uncertainty of the load prediction result is verified according to the deviation and coverage determination result.

[0038] To achieve the above objectives, the second embodiment of the present application proposes a data-driven demand-side load forecasting and uncertainty estimation system, including:

[0039] A load data collection module is used to collect historical data, wherein the historical data includes historical power load data and historical weather data at corresponding times;

[0040] A model input sample generation module is used to slice the historical data into time series of fixed length through a sliding window, use the sliced ​​data as the input sequence, and use the load value after multiple steps in the future as the prediction label to construct a training sample set;

[0041] A Bayesian deep learning module is used to construct a Bayesian deep learning load forecasting model with distribution parameters based on the ProbSparse Transformer model, optimize the model parameters of the Bayesian deep learning load forecasting model using a variational inference method according to the training sample set, and obtain a target model after the training is completed;

[0042] An uncertainty assessment module, which outputs a load forecast result according to the target model, and estimates the uncertainty and confidence interval of the load forecast result through the target model;

[0043] The verification module evaluates the prediction accuracy of the target model through indicator parameters, and verifies the uncertainty of the load prediction result through coverage judgment of the confidence interval.

[0044] The technical solution provided by the embodiments of the present application brings at least the following beneficial effects:

[0045] (1) Efficient data processing and feature extraction: By using a sliding window to slice historical power load and weather data and using the ProbSparse Transformer model for feature extraction, it can effectively process long time series data;

[0046] (2) Accurate load forecasting and uncertainty estimation: By performing variational inference optimization and uncertainty estimation on the model, confidence intervals for load forecasting can be generated, which improves the accuracy and credibility of the forecast results.

[0047] (3) High prediction accuracy: This method constructs a Bayesian deep learning model to optimize the load forecasting model parameters in an end-to-end manner, which can predict the load more accurately.

[0048] (4) Comprehensive model validation and evaluation: The model is validated using a variety of indicators such as root mean square error and mean absolute error. The deviation between the predicted value and the actual value and the coverage of the confidence interval are observed, which can comprehensively evaluate the prediction accuracy of the model and the effectiveness of uncertainty estimation.

[0049] (5) Practical system deployment and online intelligent prediction: The system can not only perform model verification and optimization during the training phase, but can also be deployed online for real-time online intelligent prediction of loads, which has high practicality and application value.

[0050] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0052] Figure 1 is a flow chart of a data-driven demand-side load forecasting and uncertainty estimation method according to an embodiment of the present application;

[0053] Figure 2 It is a curve diagram of the electric load of a city in one week according to an embodiment of the present application;

[0054] Figure 3 It is a cooling load curve diagram of a city in one week according to an embodiment of the present application;

[0055] Figure 4 is a heat load curve diagram of a city in one week according to an embodiment of the present application;

[0056] Figure 5 is a schematic diagram of load prediction results and confidence intervals according to an embodiment of the present application;

[0057] Figure 6 It is a block diagram of a data-driven demand-side load forecasting and uncertainty estimation system according to an embodiment of the present application. DETAILED DESCRIPTION

[0058] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0059] The following describes a data-driven demand-side load forecasting and uncertainty estimation method and system according to an embodiment of the present application with reference to the accompanying drawings.

[0060] Figure 1 is a flow chart of a data-driven demand-side load forecasting and uncertainty estimation method according to an embodiment of the present application. Figure 1 As shown, the method comprises the following steps:

[0061] Step 101, collecting historical data, the historical data including historical power load data and historical weather data at corresponding times.

[0062] In one embodiment of the present application, a load forecasting and uncertainty estimation method is implemented by collecting power load data of a city within a week and weather data at corresponding times.

[0063] It should be noted that the maximum sampling interval of historical power load data does not exceed 15 minutes, and the collection of historical weather data takes into account the factor variables that affect load energy consumption, including temperature, radiation, and humidity.

[0064] In a possible embodiment, the sampling interval of historical power load data is 5 minutes, and weather data is also collected every 5 minutes.

[0065] In a possible embodiment, the electric load curve of a city for one week, the cooling load curve of a city for one week, and the heat load curve of a city for one week are respectively as follows: Figure 2 , Figure 3 and Figure 4 shown.

[0066] Step 102, slice the historical data into time series of fixed length through a sliding window, use the sliced ​​data as the input sequence, and use the load value after multiple steps in the future as the prediction label to construct a training sample set.

[0067] In the embodiment of the present application, the power load and weather data are sliced ​​into time series of fixed length through a sliding window, and the time range of the sliding window is not less than 3 hours.

[0068] Then, the sliced ​​data is used as the input sequence, and the load value after multiple steps in the future is used as the predicted label to construct a training sample set, including:

[0069]

[0070]

[0071]

[0072] in, represents the training sample set, is the total number of training samples, is the number of time steps contained in the sliding window, Reason The time series input vector consists of time steps. Respectively represent Samples at time , , The true value of Indicates power load data or corresponding historical weather data, To predict the label, Represents the prediction step size.

[0073] In one embodiment of the present application, taking the load data sampling interval of 5 minutes as an example, the number of time steps included in the sliding window is Each sliding window contains 36 time steps (i.e. 3 hours of data). Each time it moves one time step, a sample is generated. Then the model prediction step size is set, i.e. the current input future The load value of the step is based on the short-term load forecast Take 1, 3, and 6 as examples, that is, predict the power load value in the next 1 step (5 minutes), 3 steps (15 minutes), and 6 steps (30 minutes).

[0074] By slicing historical data using sliding window technology, historical data can be efficiently utilized to provide the model with abundant training samples to learn the dynamic characteristics of time series, thereby improving the accuracy of load forecasting.

[0075] Step 103, based on the ProbSparse Transformer model, a Bayesian deep learning load forecasting model with distribution parameters is constructed, and the model parameters of the Bayesian deep learning load forecasting model are optimized using the variational inference method according to the training sample set. After the training is completed, the target model is obtained.

[0076] In the embodiment of the present application, based on the ProbSparse Transformer model, the steps of constructing a Bayesian deep learning load forecasting model with distribution as a parameter include:

[0077] Step 1, set the model parameters of the ProbSparse Transformer model, the model parameters include the number of attention heads, the multi-head self-attention mechanism, and the number of feedforward network layers. The multi-head self-attention mechanism is used to extract features from the input sequence, and the multi-head self-attention mechanism can adopt a sparse attention mechanism.

[0078] Among them, the number of attention heads determines the number of heads in the multi-head self-attention mechanism, and each head can independently focus on different parts of the sequence; the multi-head self-attention mechanism allows the model to capture information from different representation subspaces, and each attention head has an independent set of weight matrices; in each Transformer encoding layer and decoding layer, the feedforward network is used to further process and transform the output of the self-attention layer, which usually contains at least two linear layers and may also include normalization and activation functions.

[0079] In a possible embodiment, the number of attention heads can be set to 8 and the number of feed-forward network layers can be set to two.

[0080] It should be noted that ProbSparse Transformer uses a sparse attention mechanism to reduce computational complexity, especially when processing long sequences. This enables the model to efficiently process data of a large number of time steps without sacrificing performance due to computational resource limitations. Through the multi-head self-attention mechanism, the model can simultaneously focus on different positions in the sequence, thereby extracting richer features.

[0081] Step 2: Set the model architecture of the ProbSparse Transformer model. The model architecture includes an input layer, multiple Transformer encoding layers, multiple Transformer decoding layers, and an output layer. The output layer is used to output the load forecasting results. The load forecasting results include the load forecast value and the uncertainty of the load forecast value. The uncertainty is the variance and mean of the load forecast value.

[0082] Among them, the input layer is responsible for receiving the input time series data; each encoding layer in the multiple Transformer encoding layers contains a multi-head self-attention sublayer and a feedforward network sublayer, which is used to gradually extract and abstract the features of the input sequence; the output layer is used to generate load forecasting results.

[0083] It should be noted that in ProbSparse Transformer, the output layer not only predicts the load value, but also predicts the uncertainty of the predicted value, namely the mean and variance, to provide information about the prediction confidence.

[0084] In the ProbSparse Transformer model, in order to estimate the uncertainty of the prediction results, this application adopts a variational inference method. Specifically, the posterior distribution of the model's weight parameters is difficult to calculate directly. Therefore, this application introduces an approximate distribution, the goal of which is to minimize the KL divergence between the variational distribution and the true posterior distribution, thereby obtaining a better variational distribution estimate.

[0085] The optimization process of model parameters includes:

[0086]

[0087] in, is the variational distribution, is the true posterior distribution, represents the optimal parameters of the model, is the Kullback-Leibler divergence, which is used to measure the difference between the true posterior distribution and the variational distribution. is the weight parameter of the model.

[0088] The purpose of the above formula is to find the best parameters , so that The best approximation , which is achieved by minimizing To achieve this, we need to maximize the evidence lower bound.

[0089] As a possible implementation method, the maximum number of iterations is set to 1000, and the ProbSparseTransformer model is obtained as the target prediction model.

[0090] Step 104: output the load forecast result according to the target model, and estimate the uncertainty and confidence interval of the load forecast result through the target model.

[0091] In the embodiment of the present application, referring to step 101 and step 102, other sliced ​​data are input into the target model as an input sequence to obtain a load forecast result.

[0092] Furthermore, through the variational distribution Express model uncertainty, express random uncertainty through Gaussian distribution of the output layer, variational distribution Obtained via Monte Carlo sampling.

[0093] Specifically, the Monte Carlo sampling method is used to Draw multiple samples from Then use these samples to make predictions to get a set of prediction results.

[0094] Furthermore, the confidence interval is constructed by the standard deviation and mean of the load forecast value, and the expression is:

[0095]

[0096] in, is the mean of the load forecast values, is the standard deviation of the load forecast value.

[0097] In a possible embodiment, the load forecast result and confidence interval obtained based on a weekly heat load curve of a city are as follows: Figure 5 shown.

[0098] Step 105 , evaluating the prediction accuracy of the target model through the indicator parameters, and verifying the uncertainty of the load prediction result through the coverage judgment of the confidence interval.

[0099] In the embodiment of the present application, referring to step 101 and step 102, a test sample set is constructed to perform model verification and evaluation on the prediction accuracy of the target model, and indicators such as root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used to evaluate the model load prediction accuracy.

[0100] In a possible embodiment, a test sample of one day is selected for model testing. The evaluation indicators include root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). The diagnosis results of different methods are shown in Table 1.

[0101] Table 1

[0102]

[0103] As can be seen from Table 1, the predicted load values ​​are very close to the actual load values, and most of the actual load values ​​fall within the predicted confidence interval, indicating that the model has high prediction accuracy and reliable uncertainty estimation. In addition, the RMSE, MAE, and MAPE indicators also show that the model performs well in load forecasting. This implementation case verifies the effectiveness of the data-driven load forecasting and uncertainty estimation method and system.

[0104] In addition, the deviation between the load forecast value and the actual load value is calculated, and it is determined whether the confidence interval covers the actual load value. The validity of the uncertainty of the load forecast result is verified based on the deviation and coverage judgment results.

[0105] Specifically, for each prediction instance, calculate the deviation between the predicted value and the actual value. The deviation can be an absolute deviation or a relative deviation. By calculating the average deviation and standard deviation of all prediction instances, we can understand the accuracy and stability of the overall prediction. For each prediction instance, check whether the actual value falls within the predicted confidence interval, and count the proportion of the actual values ​​of the prediction instances that do fall within the predicted confidence interval. Ideally, this proportion should be close to the confidence level.

[0106] After obtaining the deviation and coverage judgment results, the validity of the uncertainty of the load forecast results is verified: if the average deviation is low and the standard deviation is controllable, this indicates that the model's prediction central trend is more accurate and the uncertainty estimate is relatively reasonable; if the actual coverage of the confidence interval is close to the set confidence level, this indicates that the model's estimate of the forecast uncertainty is valid. Too high or too low coverage may mean that the model underestimates or overestimates the forecast uncertainty.

[0107] In order to implement the above embodiments, the present application also proposes a data-driven demand-side load forecasting and uncertainty estimation system.

[0108] Figure 6 1 is a block diagram of a data-driven demand-side load forecasting and uncertainty estimation system 10 according to an embodiment of the present application, including:

[0109] The load data collection module 100 is used to collect historical data, including historical power load data and historical weather data at corresponding moments;

[0110] The model input sample generation module 200 is used to slice the historical data into time series of fixed length through a sliding window, use the sliced ​​data as the input sequence, and use the load value after multiple steps in the future as the prediction label to construct a training sample set;

[0111] The Bayesian deep learning module 300 is used to construct a Bayesian deep learning load forecasting model with distribution parameters based on the ProbSparse Transformer model, optimize the model parameters of the Bayesian deep learning load forecasting model using a variational inference method according to a training sample set, and obtain a target model after the training is completed;

[0112] The uncertainty assessment module 400 outputs the load forecast result according to the target model, and estimates the uncertainty and confidence interval of the load forecast result through the target model;

[0113] The verification module 500 evaluates the prediction accuracy of the target model through the indicator parameters, and verifies the uncertainty of the load prediction result through the coverage judgment of the confidence interval.

[0114] Regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0115] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this application can be executed in parallel, sequentially or in different orders, as long as the expected results of the technical solution of this application can be achieved, and this document is not limited here.

[0116] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A data-driven demand-side load forecasting and uncertainty estimation method, It is characterized in that include: Collecting historical data, wherein the historical data includes historical power load data and historical weather data at corresponding times; Slice the historical data into time series of fixed length using a sliding window, use the sliced ​​data as the input sequence, and use the load value after multiple steps in the future as the prediction label to construct a training sample set; Based on the ProbSparse Transformer model, a Bayesian deep learning load forecasting model with distribution parameters is constructed, and the model parameters of the Bayesian deep learning load forecasting model are optimized by using a variational inference method according to the training sample set. After the training is completed, a target model is obtained; Outputting a load forecast result according to the target model, and estimating the uncertainty and confidence interval of the load forecast result through the target model; The prediction accuracy of the target model is evaluated by indicator parameters, and the uncertainty of the load prediction result is tested by coverage judgment of the confidence interval; Based on the ProbSparse Transformer model, a Bayesian deep learning load forecasting model with distribution parameters is constructed, including: Set the model parameters of the ProbSparse Transformer model, including the number of attention heads, the multi-head self-attention mechanism, and the number of feedforward network layers. The multi-head self-attention mechanism is used to extract features from the input sequence. The multi-head self-attention mechanism can adopt a sparse attention mechanism. A model architecture of a ProbSparse Transformer model is set, wherein the model architecture includes an input layer, multiple Transformer encoding layers, multiple Transformer decoding layers, and an output layer, wherein the output layer is used to output a load forecast result, wherein the load forecast result includes a load forecast value and an uncertainty of the load forecast value, wherein the uncertainty is a variance and a mean of the load forecast value; The optimization process of model parameters includes: in, is the variational distribution, is the true posterior distribution, represents the optimal parameters of the model, is the Kullback-Leibler divergence, which is used to measure the difference between the true posterior distribution and the variational distribution. is the weight parameter of the model; Outputting the load forecast result according to the target model includes: Inputting other sliced ​​data as input sequences into the target model to obtain load forecasting results; The uncertainty and confidence interval of the load forecast result estimated by the target model include: Through the variational distribution Expressing model uncertainty, the variational distribution Obtained through Monte Carlo sampling; Express random uncertainty through Gaussian distribution of the output layer; The confidence interval is constructed by the standard deviation and mean of the load forecast value, and the expression is: in, is the mean of the load forecast values, is the standard deviation of the load forecast value.

2. The method according to claim 1, It is characterized in that The historical data collected includes historical power load data and historical weather data at corresponding times, including: The sampling interval of the historical power load data shall not exceed 15 minutes; The collection of the historical weather data takes into account the factor variables that affect the load energy consumption, and the factor variables include temperature, radiation, and humidity.

3. The method according to claim 2, It is characterized in that The time range of the sliding window is not less than 3 hours.

4. The method according to claim 3, It is characterized in that The sliced ​​data is used as the input sequence, and the load value after multiple steps in the future is used as the prediction label to construct a training sample set, including: in, represents the training sample set, is the total number of training samples, is the number of time steps contained in the sliding window, Reason The time series input vector consists of time steps. Respectively represent Samples at time , , The true value of Indicates power load data or corresponding historical weather data, To predict the label, Represents the prediction step size.

5. The method according to claim 1, It is characterized in that The step of evaluating the prediction accuracy of the target model by using the indicator parameters and testing the uncertainty of the load prediction result by using the coverage judgment of the confidence interval includes: Select a test sample set to perform model verification and evaluation on the prediction accuracy of the target model, and the indicator parameters include root mean square error, mean absolute error, and mean absolute percentage error; The deviation between the load prediction value and the actual load value is calculated, and it is determined whether the confidence interval covers the actual load value. The uncertainty of the load prediction result is verified according to the deviation and coverage determination result.

6. A data-driven demand-side load forecasting and uncertainty estimation system, It is characterized in that include: A load data collection module is used to collect historical data, wherein the historical data includes historical power load data and historical weather data at corresponding times; A model input sample generation module is used to slice the historical data into time series of fixed length through a sliding window, use the sliced ​​data as the input sequence, and use the load value after multiple steps in the future as the prediction label to construct a training sample set; A Bayesian deep learning module is used to construct a Bayesian deep learning load forecasting model with distribution parameters based on the ProbSparse Transformer model, optimize the model parameters of the Bayesian deep learning load forecasting model using a variational inference method according to the training sample set, and obtain a target model after the training is completed; An uncertainty assessment module, which outputs a load forecast result according to the target model, and estimates the uncertainty and confidence interval of the load forecast result through the target model; A verification module, which evaluates the prediction accuracy of the target model through indicator parameters, and verifies the uncertainty of the load prediction result through coverage judgment of the confidence interval; Based on the ProbSparse Transformer model, a Bayesian deep learning load forecasting model with distribution parameters is constructed, including: Set the model parameters of the ProbSparse Transformer model, including the number of attention heads, the multi-head self-attention mechanism, and the number of feedforward network layers. The multi-head self-attention mechanism is used to extract features from the input sequence. The multi-head self-attention mechanism can adopt a sparse attention mechanism. A model architecture of a ProbSparse Transformer model is set, wherein the model architecture includes an input layer, multiple Transformer encoding layers, multiple Transformer decoding layers, and an output layer, wherein the output layer is used to output a load forecast result, wherein the load forecast result includes a load forecast value and an uncertainty of the load forecast value, wherein the uncertainty is a variance and a mean of the load forecast value; The optimization process of model parameters includes: in, is the variational distribution, is the true posterior distribution, represents the optimal parameters of the model, is the Kullback-Leibler divergence, which is used to measure the difference between the true posterior distribution and the variational distribution. is the weight parameter of the model; Outputting the load forecast result according to the target model includes: Inputting other sliced ​​data as input sequences into the target model to obtain load forecasting results; The uncertainty and confidence interval of the load forecast result estimated by the target model include: Through the variational distribution Expressing model uncertainty, the variational distribution Obtained through Monte Carlo sampling; Stochastic uncertainty is expressed through Gaussian distribution of the output layer; The confidence interval is constructed by the standard deviation and mean of the load forecast value, and the expression is: in, is the mean of the load forecast values, is the standard deviation of the load forecast value.