Load demand prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a load demand prediction method and apparatus, an electronic device and a storage medium. The method comprises the steps of obtaining first load data; performing time sequence decomposition on the first load data to obtain seasonal data, trend data and residual data; carrying out prediction sample extraction based on seasonal data, trend data and residual data; inputting the prediction sample into a first model to obtain first prediction data; performing Pearson chi-square test calculation according to the prediction sample and the first prediction data to obtain a significance level value; when the significance level is greater than a first preset threshold value, determining that the final output is first prediction data; and when the significance level is not greater than a first preset threshold value, inputting the prediction sample into a second model to obtain second prediction data, and determining the final output as the second prediction data. Therefore, non-linear trends such as seasonality and residual errors are considered during subsequent load demand prediction.

Description

Technical Field

[0001] The present application relates to the technical field of general data processing, and particularly relates to a load demand prediction method, device, electronic device, and storage medium. Background Art

[0002] With the deepening of the power system reform, more and more user-side resources, such as distributed photovoltaic, wind power, and demand-side resources, are incorporated into the market for power trading. In this context, user-side load forecasting becomes particularly crucial because it can not only improve market participation, help the owners and operators of user-side resources participate in the market more effectively, but also promote the balance between power grid supply and demand, optimize the operation of the power grid, support the implementation of demand response strategies, improve energy efficiency, and reduce operating costs. Currently, load forecasting is mainly based on the linear trend in time series and it is difficult to take into account non-linear trends. Summary of the Invention

[0003] The present application provides a load demand prediction method, device, electronic device, and storage medium. By performing time series decomposition on load data, seasonal, trend, and residual data can be obtained, enabling subsequent load demand forecasting to take into account non-linear trends such as seasonality and residuals.

[0004] In a first aspect, the present application provides a load demand prediction method, which includes:

[0005] Obtain first load data, where the first load data is aggregated from a plurality of second load data, and the plurality of second load data are respectively historical load demand data based on time series corresponding to a plurality of users;

[0006] Perform time series decomposition on the first load data to obtain seasonal data, trend data, and residual data;

[0007] Extract prediction samples based on the seasonal data, trend data, and residual data;

[0008] Input the prediction samples into a first model to obtain first prediction data, where the first model is an autoregressive moving average-generalized autoregressive conditional heteroskedasticity model;

[0009] Perform Pearson chi-square test calculation based on the prediction samples and the first prediction data to obtain a significance level value;

[0010] When the significance level is greater than a first preset threshold, determine the final output as the first prediction data;

[0011] When the significance level is not greater than the first preset threshold, input the prediction samples into a second model to obtain second prediction data, and determine the final output as the second prediction data, where the second model is a kernel density estimation model.

[0012] It can be seen that in the present application, by performing time series decomposition on the load data, seasonal, trend, and residual data can be obtained, enabling consideration of non-linear trends such as seasonality and residuals in subsequent load demand forecasting. At the same time, when the load data distribution predicted based on the autoregressive moving average-generalized autoregressive conditional heteroskedasticity (ARMA-GARCH) model is consistent with the actual load data distribution, the prediction result of the ARMA-GARCH model is selected. When the load data distribution predicted based on the ARMA-GARCH model is inconsistent with the actual load data distribution, load forecasting is then performed through the kernel density estimation model, which can avoid probability prediction deviations caused by model misspecification.

[0013] In a feasible example, before obtaining the first load data, the method further includes: obtaining a plurality of fourth load data; assigning a random number to each of the plurality of fourth load data, and the plurality of random numbers corresponding to the plurality of fourth load data are uniformly distributed; determining a plurality of second load data from the plurality of fourth load data, and the random number corresponding to each second load data among the plurality of second load data is greater than a second preset threshold.

[0014] In the present application, by collecting original historical load demand data (i.e., fourth load data) from multiple household users or other individual users, and assigning a random number that follows a uniform distribution to each fourth load data, and screening out eligible second load data in combination with a second preset threshold, effective screening and randomization processing of user load data are achieved. This method not only enhances the flexibility of the aggregation method, but also significantly improves the quality and representativeness of the aggregated samples, reducing the impact of single-user behavior on the overall prediction result. The finally formed aggregated samples can better reflect the demand characteristics of low-aggregation-level loads, providing higher-quality input data for the subsequent ARMA-GARCH model and KDE model, thereby improving the robustness and accuracy of the prediction model.

[0015] In a feasible example, before obtaining the first load data, a plurality of fourth load data are obtained; the first load data range corresponding to the plurality of fourth load data is divided into a plurality of second load data ranges, obtaining a load data combination corresponding to each second load data range among the plurality of second load data ranges; for each load data combination among the plurality of load data combinations, a first load data combination is selected from each load data combination based on a load data selection vector, and the load data selection vector determines the selection status and non-selection status of each piece of load data in the load data combination based on a first value and a second value respectively, and the first value and the second value are not the same;

[0016] Perform time series decomposition on the load data in the first load data combination to obtain a seasonal data combination, a trend data combination, and a residual data combination; construct an objective function based on the relative standard deviation between each piece of combined data in the combined data combination and the relative standard deviation between each piece of data in the residual data combination, where the combined data combination is obtained by combining the corresponding load data between the seasonal data combination and the trend data combination; determine a first constraint condition, where the first constraint condition is used to constrain the product between the load data selection vector and the load data vector to be within a preset range; perform minimization on the objective function according to the first constraint condition to obtain the target load data selection vector; select a second load data combination from each load data combination based on the target load data selection vector; determine the load data in the multiple second load data combinations as multiple second load data.

[0017] In this application, a load demand prediction method provided in this embodiment realizes a refined description of user demand characteristics by collecting original historical load demand data (i.e., the fourth load data) from multiple household users or other individual users and dividing its total demand range into multiple sub - intervals (i.e., the second load data range); by introducing a load data selection vector, it realizes flexible screening of the load data combination, ensuring that the finally selected load data combination has higher representativeness; by performing time series decomposition on the screened load data combination, it extracts seasonal, trend, and residual data characteristics, providing a more targeted data basis for subsequent optimization; by constructing an objective function and combining the first constraint condition, it comprehensively considers the volatility of seasonal, trend, and residual data, realizing comprehensive optimization of the load data combination; finally, by performing minimization on the objective function, it obtains the optimal load data selection vector and accordingly screens out the final second load data combination, providing high - quality input data for subsequent load prediction.

[0018] In a feasible example, inputting the prediction sample into the first model to obtain the first prediction data includes: inputting the prediction sample into the first model to obtain the first prediction data and the first conditional variance; after inputting the prediction sample into the first model to obtain the first prediction data and the first conditional variance, the method further includes: generating a first probability density function according to the first prediction data and the first conditional variance; determining a first cumulative distribution function according to the first probability density function.

[0019] In this application, generating a first probability density function based on the first prediction data and the first conditional variance can more comprehensively describe the distribution characteristics of the load prediction value; subsequently, by integrating the first probability density function to obtain the first cumulative distribution function, it significantly enhances the interpretability and reliability of the load prediction result, providing more powerful support for market participants to optimize trading strategies.

[0020] In a feasible example, the prediction sample is input into the second model to obtain second prediction data, including: inputting the prediction sample into the second model to obtain second prediction data and a second probability density function; after inputting the prediction sample into the second model to obtain the second prediction data and the second probability density function, the method further includes: determining a second cumulative distribution function according to the second probability density function.

[0021] In this application, not only can the probability distribution characteristics of the second prediction data be obtained, but also the distribution range and possibility of the second prediction data can be evaluated more intuitively, providing strong support for market participants to optimize trading strategies.

[0022] In a feasible example, after determining the first cumulative distribution function according to the first probability density function, the method further includes: obtaining first real data corresponding to the first prediction data; constructing a unit step function according to the first cumulative distribution function and the first real data to obtain a first continuous ranked probability score; determining the prediction accuracy of the first cumulative distribution function according to the first continuous ranked probability score. The higher the first continuous ranked probability score, the lower the prediction accuracy of the first cumulative distribution function, and the lower the first continuous ranked probability score, the higher the prediction accuracy of the first cumulative distribution function.

[0023] In this application, the prediction performance of the first cumulative distribution function can be intuitively reflected by the level of the first continuous ranked probability score, providing an important reference basis for model optimization and selection.

[0024] In a feasible example, after obtaining the first prediction data, the method further includes: obtaining first real data corresponding to the first prediction data; determining a first mean absolute error between the first prediction data and the first real data; determining the accuracy of the first prediction data according to the magnitude of the first mean absolute error. The higher the first mean absolute error, the lower the accuracy of the first prediction data, and the lower the first mean absolute error, the higher the accuracy of the first prediction data.

[0025] In this application, by obtaining the first real data corresponding to the first prediction data, an objective reference basis is provided for the evaluation of the prediction result, avoiding the deviation that may be caused by relying only on the predicted value; by calculating the mean absolute error (MAE), a quantitative evaluation of the deviation between the predicted value and the actual observed value is realized, enhancing the scientificity and interpretability of the evaluation result; based on the level of the MAE value, the accuracy of the first prediction data is intuitively reflected, providing a clear direction for model optimization and selection.

[0026] In a feasible example, the following method can be adopted for extracting prediction samples based on seasonal data, trend data, and residual data: weighted fusion is performed on the seasonal data, trend data, and residual data to obtain third load data; prediction samples are extracted based on the third load data.

[0027] In the present application, in this way, the characteristics among the seasonal data, trend data, and residual data can be combined, thereby improving the accuracy of subsequent load demand prediction.

[0028] In a feasible example, the following method can be adopted for extracting prediction samples based on seasonal data, trend data, and residual data: the seasonal data, trend data, and residual data are concatenated to obtain a three-dimensional vector, and the three-dimensional vector is analyzed based on the self-attention mechanism to obtain three groups of feature data. The first feature data in the three groups of feature data is the feature obtained after determining the weight sizes of the query and key based on the seasonal data and trend data respectively. The second feature data in the three groups of feature data is used to represent the feature obtained after determining the weight sizes of the query and key based on the seasonal data and residual data. The third feature data in the three groups of data is used to represent the feature obtained after determining the weight sizes of the query and key based on the residual data and trend data. Weighted fusion is performed on the three groups of feature data to obtain fourth load data, and prediction samples are extracted based on the fourth load data.

[0029] In the present application, in this way, through the parameter learning of the multi-head attention and the attention mechanism, the model can explicitly model the coupling relationships between seasonality and trend, seasonality and residual, and trend and residual, so that the obtained prediction samples are more suitable for load demand prediction.

[0030] In a second aspect, the present application provides a load demand prediction device, which includes:

[0031] An acquisition unit, configured to acquire first load data, where the first load data is aggregated according to a plurality of second load data, and the plurality of second load data are respectively historical load demand data based on time series corresponding to a plurality of users;

[0032] A processing unit, configured to perform time series decomposition on the first load data to obtain seasonal data, trend data, and residual data;

[0033] The processing unit is further configured to extract prediction samples based on the seasonal data, trend data, and residual data;

[0034] The processing unit is further configured to input the prediction samples into a first model to obtain first prediction data, where the first model is an autoregressive moving average-generalized autoregressive conditional heteroskedasticity model;

[0035] The processing unit is further configured to perform a Pearson chi-square test calculation based on the prediction sample and the first prediction data to obtain a significance level value;

[0036] The processing unit is further configured to determine that the final output is the first prediction data when the significance level is greater than the first preset threshold;

[0037] The processing unit is further configured to, when the significance level is not greater than the first preset threshold, input the prediction sample into the second model to obtain second prediction data, and determine that the final output is the second prediction data, where the second model is a kernel density estimation model.

[0038] In a third aspect, the present application provides an electronic device, which includes a processor, a memory, and a communication interface. The processor, the memory, and the communication interface are interconnected and complete communication with each other. The memory stores executable program code. The communication interface is used for wireless communication. The processor is used to retrieve the executable program code stored on the memory and execute some or all of the steps described in any method of the first aspect.

[0039] In a fourth aspect, the present application provides a computer-readable storage medium, in which electronic data is stored. When the electronic data is executed by a processor, it is used to execute the electronic data to implement some or all of the steps described in the first aspect of the present application.

[0040] In a fifth aspect, the present application provides a computer program product, including a computer program, which can be operated to cause a computer to execute some or all of the steps described in the first aspect of the present application. The computer program product can be a software installation package. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 It is a schematic structural diagram of a load demand prediction system provided by an embodiment of the present application;

[0043] Figure 2 It is a schematic flowchart of a load demand prediction method provided by an embodiment of the present application;

[0044] Figure 3 It is a schematic flowchart of another load demand prediction method provided by an embodiment of the present application;

[0045] Figure 4 Schematic flowchart of another load demand prediction method provided by an embodiment of the present application;

[0046] Figure 5 Schematic flowchart of another load demand prediction method provided by an embodiment of the present application;

[0047] Figure 6 Block diagram of the functional units of a load demand prediction device provided by an embodiment of the present application;

[0048] Figure 7 Block diagram of the functional units of another load demand prediction device provided by an embodiment of the present application;

[0049] Figure 8 Block diagram of the structure of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0050] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0051] Terms such as "first" and "second" in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps is not limited to the listed steps, but optionally further includes steps not listed, or optionally further includes other steps inherent to these processes, methods, products or devices.

[0052] Referring to "embodiment" herein means that a specific feature, structure or characteristic described in connection with the embodiment may be included in at least one embodiment of the present application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein may be combined with other embodiments.

[0053] The load curves of medium and large industrial and commercial users are relatively stable with small fluctuations and are less affected by factors such as user behavior habits and seasonal changes. Therefore, the difficulty of load forecasting for them is relatively low. However, for individual users with small demand, large volatility, large numbers, and wide distribution (such as household users), the load curve has high randomness and obvious peak-valley changes, making the load forecasting difficult. To solve this problem, load aggregators or virtual power plant operators are usually used to aggregate the scattered individual user loads together to form a low-aggregation-level load for market trading. Through accurate low-aggregation-level load forecasting, individual household users can reasonably arrange market trading strategies and are conducive to improving the operation of point-to-point energy sharing, demand response, and reliable distribution networks. Therefore, how to make full use of the forecasting model to exert the accurate short-term forecasting ability of low-aggregation-level loads, so that various market participants can optimize trading strategies, is the key to solving problems such as market complexity and grid stability brought about by the participation of user-side resources in market trading.

[0054] Short-term low-aggregation-level load forecasting usually exists in decentralized or de-centralized energy markets, which include multiple market participants such as individual consumers, small producers, demand response participants, and grid operators. Load forecasting is based on factors such as historical load data, weather information, time series analysis, user behavior patterns, and possible external events, and uses statistical models, machine learning, and artificial intelligence algorithms to analyze this data to predict the load changes in the short term in the future. By providing accurate short-term load forecasting for market participants, it can help market players make more informed buying and selling decisions and optimize energy trading strategies; at the same time, it can promote market participants to adjust their own electricity consumption behaviors according to the forecasting results to respond to grid demand changes and participate in demand response programs; for grid operators, load forecasting can optimize the allocation and scheduling of power resources, maintain the reliability and stability of the grid, help users improve energy use efficiency, reduce waste, and lower energy costs. Generally speaking, short-term load forecasting usually requires the integration of advanced information technologies such as big data analysis, cloud computing, Internet of Things (IoT) devices, and smart meters to achieve real-time data collection and processing.

[0055] In general, most studies are limited to system-level loads or individual household loads that rely on smart meter data. However, in the face of the challenges of individual households in peer-to-peer energy trading and demand-side response, such as insufficient bargaining power due to small electricity loads, the industry has started to aggregate the loads of multiple households through load aggregators or virtual power plants. In this case, the load aggregator or virtual power plant aggregates the loads of different household users to participate in market transactions to optimize energy distribution and reduce transaction costs. At this time, predicting the aggregated load is more representative. Among them, LSTM+RNN (Long Short-Term Memory Network combined with Recurrent Neural Network) is a deep learning technology commonly used for short-term load forecasting at a low aggregation level. LSTM is a special type of RNN (Recurrent Neural Network) that can learn long-term dependencies. It solves the problem of gradient disappearance in traditional RNNs by introducing three gates (input gate, forget gate, output gate).

[0056] Under the construction path of the new power system, with the continuous deepening of the participation of user-side resources in market transactions, the complexity of market transactions has increased significantly. Considering the diversity of user behavior and the popularity of distributed energy on the user side, the difficulty of load forecasting at a low aggregation level has increased accordingly. Compared with the initial stage of the construction of the power market, the gradual liberalization of the user-side distributed power market at the present stage puts forward higher requirements for the forecasting ability of the aggregated user-side load. Existing load forecasting mainly considers the linear trend characteristics in historical load data and does not consider the non-linear trend characteristics in historical load data.

[0057] Based on this, the present application provides a load demand forecasting method. For the aggregated load data, time series decomposition is performed to obtain seasonal data, trend data, and residual data. Then, prediction samples are extracted based on the foregoing three types of data, and the prediction samples are input into the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model for load forecasting. At the same time, when the distribution of the load data predicted by the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model is consistent with the distribution of the actual load data, the prediction result of the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model is selected. When the distribution of the load data predicted by the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model is inconsistent with the distribution of the actual load data, load forecasting is performed again through the kernel density estimation model, which can avoid the probability prediction deviation caused by model misspecification.

[0058] Please refer to Figure 1 , Figure 1 which is a schematic structural diagram of a load demand forecasting system provided by an embodiment of the present application. As Figure 1 shown, the load demand forecasting system 100 includes a load data acquisition module 101 and a load demand forecasting module 102.

[0059] The load data acquisition module 101 is used to acquire the load data of multiple household users and transmit it to the load demand prediction module 102. The load data acquisition module 101 can be a smart meter.

[0060] The load demand prediction module 102 is used to predict the load demand based on the load data of multiple household users. The load demand prediction module 102 can be a terminal device, such as a desktop computer, a laptop computer, a tablet computer, a smart phone, etc. It can also be a server, such as a single server, a server cluster, a cloud server, a cloud computing service center, or other forms of devices with computing capabilities.

[0061] Specifically, after the load demand prediction module 102 obtains the first load data from the load data acquisition module 101, it will perform time series decomposition on the first load data to obtain seasonal data, trend data, and residual data. And extract prediction samples based on at least one of the seasonal data, trend data, and residual data. Subsequently, the prediction samples are input into an Autoregressive Moving Average-Generalized Autoregressive Conditional Heteroskedasticity (ARMA-GARCH) model to obtain the first prediction data; perform Pearson chi-square test calculation based on the prediction samples and the first prediction data to obtain a significance level value; when the significance level is greater than the first preset threshold, determine that the final output is the first prediction data; when the significance level is not greater than the first preset threshold, input the prediction samples into a Kernel Density Estimation (KDE) model to obtain the second prediction data, and determine that the final output is the second prediction data.

[0062] Based on the time series decomposition of the load data to obtain seasonal, trend, and residual data, it can enable subsequent load demand prediction to consider non-linear trends such as seasonality and residuals. At the same time, when the load data distribution predicted by the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model is consistent with the actual load data distribution, select the prediction result of the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model. When the load data distribution predicted by the autoregressive moving average-generalized autoregressive conditional heteroskedasticity model is inconsistent with the actual load data distribution, perform load prediction through the kernel density estimation model again, which can avoid probability prediction deviations caused by model mis-specification.

[0063] Based on this, an embodiment of the present application provides a load demand prediction method, which will be described in detail below in conjunction with the accompanying drawings.

[0064] Example 1. Next, the framework of the load demand forecasting method in the embodiments of the present application will be described.

[0065] Please refer to Figure 2 , Figure 2 , which is a schematic flowchart of a load demand forecasting method provided by an embodiment of the present application. This method is applied to the above-mentioned load demand forecasting module, as Figure 2 shown. The method includes the following steps:

[0066] Step S201: Obtain the first load data.

[0067] Among them, the first load data is aggregated from multiple second load data, and the multiple second load data are respectively historical load demand data based on time series corresponding to multiple users. The second load data can be collected based on smart meters. Exemplarily, the second load data may include, but is not limited to, load demand data per hour, per day, or per week.

[0068] Optionally, before obtaining the first load data, obtain multiple fourth load data; assign a random number to each of the multiple fourth load data, and the multiple random numbers corresponding to the multiple fourth load data are uniformly distributed; determine multiple second load data from the multiple fourth load data, and the random number corresponding to each second load data among the multiple second load data is greater than a second preset threshold.

[0069] Among them, each fourth load data can be the original historical load demand data obtained from a single household user or other individual users, and can be collected through smart meters or other data collection devices. Exemplarily, the fourth load data may include, but is not limited to, information such as the electricity consumption of users and the distribution of electricity usage time. These data can reflect the behavior patterns and electricity usage habits of different users, thus providing diverse inputs for subsequent aggregation.

[0070] Subsequently, assign a random number to each of the multiple fourth load data to ensure that the multiple random numbers corresponding to the multiple fourth load data are uniformly distributed. The random number can be a value that follows a uniform distribution and is used to determine whether a certain user is selected into the aggregation sample.

[0071] Further, a plurality of second load data is determined from the plurality of fourth load data. Among them, the second load data can be screened out from the fourth load data, and the screening condition is that the random number corresponding to each fourth load data is greater than a second preset threshold. The second preset threshold can be a set critical value for screening eligible users. Exemplarily, the second load data can include, but is not limited to, user load data meeting specific conditions. By setting the second preset threshold, the proportion of users entering the aggregated sample can be controlled, which not only ensures the representativeness of the aggregated sample but also avoids too many or too few users participating in the aggregation, thereby improving the accuracy of subsequent predictions.

[0072] Exemplarily, for each household user i, a random number r is generated i , where r i follows a uniform distribution, that is, r i ∽U(0,1). Define a threshold set where k represents the number of aggregation levels to be described. For each threshold θ ∈ Θ, construct a portfolio V(θ). If the random number r of household i i is less than the threshold θ, then this household is included in V(θ). Whether household i is included in the portfolio can be represented by an indicator function I:

[0073]

[0074] where: I i (θ) being 1 indicates that household i is selected, and being 0 indicates not being selected. It can be understood that when household i is selected, the load data corresponding to this household i is determined as the second load data.

[0075] In this application, by collecting the original historical load demand data (i.e., the fourth load data) from multiple household users or other individual users, and assigning a random number following a uniform distribution to each fourth load data, the eligible second load data is screened out in combination with the second preset threshold, thus realizing the effective screening and randomization processing of user load data. This method not only enhances the flexibility of the aggregation method but also significantly improves the quality and representativeness of the aggregated sample, reducing the impact of single-user behavior on the overall prediction result. The finally formed aggregated sample can better reflect the demand characteristics of low-aggregation-level loads, providing higher-quality input data for the subsequent ARMA-GARCH model and KDE model, thereby improving the robustness and accuracy of the prediction model.

[0076] Optionally, before obtaining the first load data, obtain a plurality of fourth load data; divide the first load data range corresponding to the plurality of fourth load data into a plurality of second load data ranges to obtain a load data combination corresponding to each second load data range in the plurality of second load data ranges; for each load data combination in the plurality of load data combinations, select a first load data combination from each load data combination based on a load data selection vector, where the load data selection vector determines the selection status and non-selection status of each load data in the load data combination based on a first value and a second value respectively, and the first value and the second value are different;

[0077] Perform time series decomposition on the load data in the first load data combination to obtain a seasonal data combination, a trend data combination, and a residual data combination; construct an objective function based on the relative standard deviation between each merged data in the merged data combination and the relative standard deviation between each data in the residual data combination, where the merged data combination is obtained by merging the load data corresponding to the seasonal data combination and the trend data combination; determine a first constraint condition, where the first constraint condition is used to constrain the product between the load data selection vector and the load data vector to be within a preset range; perform minimization solution on the objective function according to the first constraint condition to obtain a target load data selection vector; select a second load data combination from each load data combination based on the target load data selection vector; determine the load data in the plurality of second load data combinations as the plurality of second load data.

[0078] Among them, the first load data range can be the total demand range of all the fourth load data, which can be obtained by statistically calculating the maximum value and the minimum value of the fourth load data. Exemplarily, the first load data range can be expressed as [0, 100]. The second load data range can be the result of dividing the first load data range into a plurality of sub-intervals. Exemplarily, [0, 100] can be divided into K equal-sized partitions, such as [0, 25], [25, 50],..., [75, 100]. In a specific embodiment, by dividing the first load data range of all the fourth load data into a plurality of second load data ranges, with each range corresponding to a possible load data combination, the demand characteristics of different user groups can be described more precisely, thereby providing a more specific partitioning basis for subsequent portfolio optimization.

[0079] The load data selection vector can be used to determine the selection status (selected or unselected) of each piece of load data in the load data combination, which can be achieved by defining the element values in the vector as the first value or the second value. The first value and the second value can respectively represent the flag values of the selected state and the unselected state, and the two are different. Exemplarily, the first value can be 1 and the second value can be 0. In a specific embodiment, for each load data combination, the selection status of each piece of load data is determined based on the load data selection vector, so as to screen out the first load data combination that meets the conditions. By introducing the load data selection vector, flexible screening of the load data combination is realized, ensuring that the finally selected load data combination has higher representativeness.

[0080] Time series decomposition can be a process of decomposing load data into seasonal data, trend data, and residual data, which can also be achieved by the STL method. Exemplarily, the seasonal data can reflect the periodic fluctuation characteristics, the trend data can reflect the long-term change trend, and the residual data can reflect the random noise. In a specific embodiment, the STL method is used to perform time series decomposition on the first load data combination to extract the seasonal data combination, the trend data combination, and the residual data combination. Through time series decomposition, the composition components of the load data can be understood more clearly, providing a more targeted data basis for the subsequent construction of the objective function.

[0081] The combined data combination can be a data set obtained by combining the seasonal data combination and the trend data combination. The relative standard deviation can be an index to measure the data volatility, and the calculation formula is the sample standard deviation divided by its average value. Exemplarily, the combined data combination can include the comprehensive characteristics of seasonal and trend data, and the relative standard deviation can quantify the stability of these data. In a specific embodiment, an objective function is constructed based on the relative standard deviation between each piece of combined data in the combined data combination and the relative standard deviation between each piece of data in the residual data combination. The goal is to minimize the objective function value by adjusting the load data selection vector. By constructing the objective function, the volatility of seasonal, trend, and residual data can be comprehensively considered, thereby realizing the optimal selection of the load data combination.

[0082] The first constraint condition can be used to limit the product of the load data selection vector and the load data vector to be within a preset range, which can be achieved by defining the upper and lower limits of the constraint condition. In a specific embodiment, the first constraint condition is defined to ensure that the product of the load data selection vector and the load data vector falls within the preset range. Through the first constraint condition, it is ensured that the finally selected load data combination is within the actual demand range and avoids deviating too much from the actual situation.

[0083] The target load data selection vector can be the optimal load data selection vector obtained by minimizing the objective function, which can be achieved through a genetic algorithm or other optimization algorithms. Exemplarily, the genetic algorithm can find the optimal solution by simulating the natural selection process. In a specific embodiment, in combination with the first constraint condition, the objective function is minimized to obtain the target load data selection vector. By minimizing the objective function, the optimal load data selection vector can be found, thereby achieving precise optimization of the load data combination.

[0084] The second load data combination can be the final load data set screened from each load data combination based on the target load data selection vector. Exemplarily, the second load data combination can include an optimized load data subset. In a specific embodiment, according to the target load data selection vector, the second load data combination that meets the conditions is selected from each load data combination. Through screening, it is ensured that the finally selected second load data combination has higher representativeness and prediction accuracy.

[0085] The multiple second load data can be the result obtained by summarizing the load data in all the second load data combinations. Exemplarily, the multiple second load data can include optimized load data sets from different partitions. In a specific embodiment, by summarizing the load data in all the second load data combinations, the final multiple second load data is formed. Through summarization, the final load data set is generated, providing high-quality input data for subsequent load forecasting.

[0086] Exemplarily, let Y be a (T×N) matrix containing T demand observations and N household users; at time node t, the load data range [D min , D max of the N household users can be divided into K equally sized partitions to obtain the optimal portfolio for each partition k.

[0087] Among them, the load data range corresponding to the kth partition is: [D min +(k - 1)·ΔD, D min +k·ΔD], where k = 1, 2..., K; ΔD = (D max - D min ) / K.

[0088] Therefore, the corresponding objective function g can be:

[0089]

[0090] The first constraint condition is:

[0091]

[0092] Among them, Y represents the load data matrix for household users; represents the load data vector of N household users in the prediction period h; is an (N×1) vector, which can include 1 or 0, indicating whether the household users in area k are selected respectively; c (k,h) and c (k+1,h) respectively represent the minimum load data value and the maximum load data value of the k-th partition at the prediction period h. The objective function aims to minimize the prediction error by changing parameters such as the vector, for example the vector, thus the vector is the finally selected household user portfolio selection, which can minimize the probabilistic prediction error at a specific prediction period h and the load data of the k-th partition of the objective function g.

[0093] The genetic algorithm is used to solve the objective function. For the objective function, the seasonal similarity method (SS) is adopted, that is, by minimizing the deviation of the seasonal signals of the household users in the portfolio. The SS method is a bi-objective optimization method using the target ratio r, where SS uses a subset (matrix k) of the load demand observation matrix Y for demand prediction. The specific formula is as follows:

[0094]

[0095] Among them, the matrix and are respectively the corresponding seasonal, trend and residual time series under the household user selection vector in area k; refers to the time series decomposition method of the load demand observation matrix under the household user selection vector ; the cycle length t generally takes h as the unit. The SS method minimizes the following formula by selecting the portfolio with the smallest change in the standard deviation of the seasonal, trend (these two parts as objective 1) and residual (this part as objective 2) parts, where the bi-objective weight ratio is controlled by the parameter r, thus obtaining the following formula:

[0096]

[0097] Among them, represents the relative standard deviation, and the calculation formula is the sample standard deviation divided by its average value; by changing the vector in the sample and the weight r to minimize the objective function where the weight r is allowed to change with the prediction period.

[0098] In this application, a load demand forecasting method provided by this embodiment realizes a refined description of user demand characteristics by collecting original historical load demand data (i.e., the fourth load data) from multiple household users or other individual users and dividing its total demand range into multiple sub-intervals (i.e., the second load data range); by introducing a load data selection vector, it realizes flexible screening of load data combinations, ensuring that the finally selected load data combination has higher representativeness; by performing time series decomposition on the screened load data combination to extract seasonal, trend, and residual data characteristics, it provides a more targeted data basis for subsequent optimization; by constructing an objective function and combining the first constraint condition, it comprehensively considers the volatility of seasonal, trend, and residual data, realizing the overall optimization of the load data combination; finally, by minimizing the objective function, the optimal load data selection vector is obtained, and based on this, the final second load data combination is screened out, providing high-quality input data for subsequent load forecasting.

[0099] Step S202: Perform time series decomposition on the first load data to obtain seasonal data, trend data, and residual data.

[0100] Among them, time series decomposition can be a process of decomposing time series data into seasonal data, trend data, and residual data, which can be implemented by using the Seasonal-Trend decomposition procedure based on Loess (STL) method. Exemplarily, the STL method can handle missing information and outliers in the time series and extract the periodic pattern (seasonality), long-term change trend (trend), and unexplained random fluctuations (residuals) in the time series. Through this operation, the composition of the load data can be understood more clearly, providing a more targeted data basis for subsequent forecasting.

[0101] The decomposition formula corresponding to the STL method is:

[0102] Y v =S v +T v +R v

[0103] Among them, S v is the seasonal data, T v is the trend data, and R v is the residual data.

[0104] Step S203: Extract prediction samples based on the seasonal data, trend data, and residual data.

[0105] Among them, it can be obtained by selecting one or more data types from the above decomposition results. In a specific embodiment, predictive samples can be extracted from seasonal data, trend data, and residual data using a sliding window method. The sliding window can be a data sampling method for extracting training sets and test sets by gradually moving a time window of a fixed length. By defining the length of the sliding window (such as 12 weeks as the training period and 72 hours as the validation period), and moving to the next time period after each training to extract new samples. Through this operation, the training data can be dynamically updated to ensure that the model is always trained and predicted based on the latest data.

[0106] Optionally, the following method can be used to extract predictive samples based on seasonal data, trend data, and residual data: perform weighted fusion on seasonal data, trend data, and residual data to obtain third load data; extract predictive samples based on the third load data.

[0107] Optionally, the following method can be used to extract predictive samples based on seasonal data, trend data, and residual data: splice seasonal data, trend data, and residual data to obtain a three-dimensional vector, analyze the three-dimensional vector based on the self-attention mechanism to obtain three sets of feature data. The first feature data in the three sets of feature data is the feature obtained after determining the weight sizes of the query and key based on seasonal data and trend data respectively. The second feature data in the three sets of feature data is used to represent the feature obtained after determining the weight sizes of the query and key based on seasonal data and residual data. The third feature data in the three sets of data is used to represent the feature obtained after determining the weight sizes of the query and key based on residual data and trend data. Perform weighted fusion on the three sets of feature data to obtain fourth load data, and extract predictive samples based on the fourth load data.

[0108] Among them, the aforementioned self-attention mechanism includes three attention heads, which are respectively used to analyze seasonal data and trend data, seasonal data and residual data, and residual data and trend data. The following is an explanation of the first attention head for analyzing seasonal data and trend data. The goal of the first attention head is to capture the coupling between seasonality (S) and trend (T). Each attention head includes a query (Q), a key (K), and a value (V). F is the three-dimensional vector obtained by splicing seasonal data, trend data, and residual data. The weight is larger in the seasonal dimension, the weight is larger in the trend dimension, All components are retained. As a result, the first feature data obtained by the first attention head is the feature obtained by determining the weights of the query and key based on the seasonal data and the trend data respectively.

[0109] In this way, through the parameter learning of the multi-head attention and the attention mechanism, the model can explicitly model the coupling relationships between seasonality and trend, seasonality and residual, and trend and residual, so that the obtained prediction samples are more suitable for load demand prediction.

[0110] Step S204, input the prediction sample into the first model to obtain the first prediction data.

[0111] Among them, the first model is an autoregressive moving average-generalized autoregressive conditional heteroskedasticity model. It can be achieved by fitting the ARMA(p,q) model to capture the autocorrelation of the conditional mean, and at the same time fitting the GARCH(r,s) model to capture the autocorrelation of the conditional variance. Exemplarily, the ARMA-GARCH model can calculate the mean and volatility of time series data, so that it can not only predict the load value, but also quantify the uncertainty of the prediction, providing more comprehensive information for market participants.

[0112] The ARMA-GARCH model is implemented based on the following formula:

[0113]

[0114] Among them, y t represents the prediction data at time t; y t-j represents the load data at time t-j judged by the random method and the investment optimization method; ε t represents the error term; σ t is the conditional standard deviation representing volatility; α, β, φ, γ represent the coefficients of the AR autoregression, MA moving average, GARCH generalized autoregressive conditional heteroskedasticity, and ARCH autoregressive conditional heteroskedasticity parts respectively; p, q are the orders of the ARMA model; r, s are the orders of the GARCH model; η t represents the generation process of white noise.

[0115] Specifically, the order of the ARMA model can be selected by the Bayesian Information Criterion (BIC). BIC is an index to measure the goodness of fit of the model, and at the same time penalizes the complexity of the model (i.e., the number of model parameters).

[0116] First, for a series of possible combinations of p and q values, ARMA models are respectively fitted. For example, considering the maximum values of p and q are 3, the model combinations to be fitted are: ARMA(0,0), ARMA(0,1),..., ARMA(1,0), ARMA(1,1),..., ARMA(3,3).

[0117] Secondly, for each fitted ARMA model, its likelihood function is calculated Likelihood function represents the probability of the occurrence of the predicted data y under given parameters. The likelihood function of the ARMA model is based on the probability density function of the residual ε t (error term). Assuming that the residuals are independently and identically distributed normal distributions, i.e., white noise, we have:

[0118] ε t ~N(0,σ 2 )

[0119] Given the predicted data y t and the model parameters θ (including the autoregressive part and the moving average part), the parameter formula for each time point is:

[0120]

[0121] Since the residuals follow a normal distribution, its probability density function formula is:

[0122]

[0123] For the entire time series, the formula for the joint probability density function (i.e., the likelihood function) is:

[0124]

[0125] Thirdly, the BIC is calculated using the likelihood function of each model, the number of model parameters k (where k = p + q), and the sample size n. The formula is as follows:

[0126]

[0127] Compare the BIC values of all models. A lower BIC value indicates that the model provides a better fit after penalty. Select the ARMA model with the lowest BIC value as the final model, which achieves the best balance between fitting the data and keeping the model simple.

[0128] Further, the GARCH model is used to describe the volatility of time series, where r and s represent the orders of the GARCH and ARCH parts respectively. The order of the GARCH model can be custom-set. For example, a common setting is (1,1). The GARCH(1,1) model is sufficient to capture the volatility clustering phenomenon of many time series data while maintaining the simplicity of the model.

[0129] Step S205: Perform Pearson chi-square test calculation based on the prediction samples and the first prediction data to obtain the significance level value.

[0130] Among them, the Pearson chi-square test can be a method for evaluating whether the distribution predicted by the model is consistent with the actual data distribution. It can be achieved by calculating the chi-square statistic and calculating the significance level value (p-value) through the cumulative distribution function (CDF) of the chi-square distribution. Exemplarily, the p-value can reflect the distribution difference between the prediction samples and the actual data. Through this operation, it can be judged whether the prediction result of the ARMA-GARCH model conforms to the actual data distribution.

[0131] The expression corresponding to the Pearson chi-square test is:

[0132]

[0133] After calculating the chi-square statistic the p-value is calculated by the cumulative distribution function (CDF) of the chi-square distribution, and the formula is as follows:

[0134]

[0135] where df = w - (p + q + r + s); W represents the total number of load observation values; y t-w represents the prediction sample; y represents the load prediction value.

[0136] Step S206: When the significance level is greater than the first preset threshold, determine that the final output is the first prediction data.

[0137] Among them, if the significance level value is greater than the first preset threshold, it is considered that the prediction distribution of the ARMA-GARCH model is consistent with the actual data distribution, and the first prediction data is directly output.

[0138] Exemplarily, judge according to the first preset threshold δ (determined according to the confidence level): when p > δ = 0.05 (95% confidence level), accept the prediction result (the first prediction data) of the ARMA-GARCH model.

[0139] Step S207: When the significance level is not greater than the first preset threshold, input the prediction sample into the second model to obtain the second prediction data, and determine that the final output is the second prediction data.

[0140] Among them, the second model is a kernel density estimation model. When the significance level value (p-value) is not greater than the first preset threshold, it is considered that the ARMA-GARCH model cannot accurately describe the data distribution. Exemplarily, it is judged according to the first preset threshold δ (determined according to the confidence level): p ≤ δ = 0.05 (95% confidence level), and the prediction result of the ARMA-GARCH model is rejected.

[0141] Exemplarily, in the case where the significance level is not greater than the first preset threshold, applying the KDE model can avoid any assumptions about the distribution of the ARMA-GARCH model and generate a density function based on historical load data. This method estimates the probability distribution function f(y) of the prediction data y according to the historical load data {y1, y2,..., y n}, and the specific expression is:

[0142]

[0143] where: y represents the prediction data; W represents the total number of load data; k h represents a Gaussian kernel function with bandwidth h. The bandwidth h needs to be fully estimated because too large a bandwidth will cause excessive smoothing of the key basic features of the time series, while too small a bandwidth will cause insufficient smoothing, resulting in a too rough predicted density.

[0144] The regularized selection expression of the bandwidth is as follows:

[0145] h = (4σ 5 / 3W) 1 / 5

[0146] σ is the standard deviation of the load data, and the specific expression is:

[0147]

[0148] where, represents the mean value of the load data.

[0149] It can be seen that a load demand prediction method provided by an embodiment of the present application aggregates the load data of multiple users to reduce randomness and volatility, and uses the time series decomposition method to decompose the time series data to extract seasonal, trend, and residual data, which can make the subsequent load demand prediction take into account non-linear trends such as seasonality and residuals. At the same time, when the load data distribution predicted based on the ARMA-GARCH model is consistent with the actual load data distribution, the prediction result of the ARMA-GARCH model is selected; when the load data distribution predicted based on the ARMA-GARCH model is inconsistent with the actual load data distribution, the KDE model is used for load prediction, which can avoid the probability prediction deviation caused by model mis-specification.

[0150] Embodiment 2. The load demand prediction method in the embodiments of the present application will be described below in combination with the evaluation method of prediction data.

[0151] Please refer to Figure 3 , Figure 3 which is a schematic flowchart of another load demand prediction method provided by the embodiments of the present application. This method is applied to the above-mentioned load demand prediction module, as Figure 3 shown. This method includes the following steps:

[0152] Step S301, obtain the first load data.

[0153] Among them, the first load data is aggregated according to multiple second load data, and the multiple second load data are respectively historical load demand data based on time series corresponding to multiple users.

[0154] Step S302, decompose the first load data by time series to obtain seasonal data, trend data, and residual data.

[0155] Step S303, extract prediction samples based on the seasonal data, trend data, and residual data.

[0156] Step S304, input the prediction samples into the first model to obtain the first prediction data.

[0157] Among them, the first model is an autoregressive moving average-generalized autoregressive conditional heteroskedasticity model.

[0158] Step S305, perform Pearson chi-square test calculation according to the prediction samples and the first prediction data to obtain the significance level value.

[0159] Step S306, when the significance level is greater than the first preset threshold, determine that the final output is the first prediction data.

[0160] Step S307, obtain the first real data corresponding to the first prediction data.

[0161] Among them, the first real data can be the actual observed value corresponding to the first prediction data, which is used to evaluate the accuracy of the prediction result. The first real data can be obtained by extracting the time point and load value corresponding to the first prediction data from the actual observed data. Exemplarily, assuming that the first prediction data is the predicted value of the load demand for a certain period, the first real data may include, but is not limited to, the actual load demand value within the same period, etc. By obtaining the first real data, a real reference basis can be provided for the subsequent prediction accuracy evaluation, so as to more objectively evaluate the performance of the model.

[0162] Step S308, determine the first mean absolute error between the first predicted data and the first actual data;

[0163] The mean absolute error is an index for evaluating prediction accuracy and is used to measure the average deviation between the predicted value and the actual observed value. In a specific embodiment, calculate the absolute error between the first predicted data and the first actual data, and take its average value as the mean absolute error (MAE). The specific formula is:

[0164]

[0165] where n represents the total number of samples, is used to represent the first actual data, y t represents the first predicted data.

[0166] Step S309, determine the accuracy of the first predicted data according to the magnitude of the first mean absolute error.

[0167] Among them, the higher the first mean absolute error, the lower the accuracy of the first predicted data, and the lower the first mean absolute error, the higher the accuracy of the first predicted data.

[0168] It can be seen that by obtaining the first actual data corresponding to the first predicted data, an objective reference basis is provided for the evaluation of the prediction result, avoiding the deviation that may be caused by relying solely on the predicted value; by calculating the mean absolute error (MAE), a quantitative evaluation of the deviation between the predicted value and the actual observed value is realized, enhancing the scientificity and interpretability of the evaluation result; based on the level of the MAE value, the accuracy of the first predicted data is intuitively reflected, providing a clear direction for model optimization and selection.

[0169] Step S310, when the significance level is not greater than the first preset threshold, input the prediction sample into the second model to obtain the second predicted data, and determine that the final output is the second predicted data.

[0170] Among them, the second model is a kernel density estimation model.

[0171] Step S311, obtain the second actual data corresponding to the second predicted data.

[0172] Step S312, determine the second mean absolute error between the second predicted data and the second actual data;

[0173] Step S313, determine the accuracy of the second predicted data according to the magnitude of the second mean absolute error.

[0174] Among them, the higher the second mean absolute error, the lower the accuracy of the second predicted data, and the lower the second mean absolute error, the higher the accuracy of the second predicted data. The accuracy evaluation of the second predicted data is similar to that of the first predicted data, which will not be elaborated here.

[0175] Embodiment 3. Next, the load demand prediction method in the embodiments of the present application will be described again in combination with the evaluation method of predicted data.

[0176] Please refer to Figure 4 , Figure 4 which is a schematic flowchart of another load demand prediction method provided by the embodiments of the present application. This method is applied to the above load demand prediction module, as Figure 4 shown. This method includes the following steps:

[0177] Step S401, obtain the first load data.

[0178] Among them, the first load data is aggregated from multiple second load data, and the multiple second load data are respectively historical load demand data based on time series corresponding to multiple users.

[0179] Step S402, perform time series decomposition on the first load data to obtain seasonal data, trend data, and residual data.

[0180] Step S403, extract prediction samples based on the seasonal data, trend data, and residual data.

[0181] Step S404, input the prediction samples into the first model to obtain the first predicted data and the first conditional variance.

[0182] Among them, the first model is an ARMA-GARCH model. The first predicted data can be the load demand prediction value generated by the ARMA model. Exemplarily, it represents the load demand prediction value at a certain future moment. The first conditional variance can be the volatility prediction value generated by the GARCH model. Exemplarily, it represents the uncertainty range of the prediction value. The first conditional variance can refer to the

[0183] Step S405, generate the first probability density function according to the first predicted data and the first conditional variance.

[0184] Among them, the first probability density function can be a probability distribution function (PDF) generated based on the first prediction data and the first conditional variance, which is used to describe the possible distribution of the load prediction value. By using the first prediction data as the mean and the first conditional variance as the variance, a probability density function of a normal distribution can be generated. Through this operation, the distribution characteristics of the prediction data can be described more comprehensively, including not only the prediction data itself but also its uncertainty range, thus providing a basis for the subsequent calculation of the cumulative distribution function (CDF).

[0185] Step S406: Determine the first cumulative distribution function according to the first probability density function.

[0186] Among them, the first cumulative distribution function can be a cumulative distribution function obtained by integrating the first probability density function, which is used to describe the probability that the load prediction value is less than or equal to a certain specific value. Exemplarily, the cumulative distribution function can be expressed by the following formula:

[0187]

[0188] Among them, f(u) represents the probability density function, u is the integration variable, which is used to accumulate the integral of the probability density function between -∞ and a certain value x; x represents the prediction data.

[0189] It can be seen that generating the first probability density function based on the first prediction data and the first conditional variance can describe the distribution characteristics of the load prediction value more comprehensively; subsequently, obtaining the first cumulative distribution function by integrating the first probability density function significantly enhances the interpretability and reliability of the load prediction result, providing stronger support for market participants to optimize trading strategies.

[0190] Step S407: Perform Pearson chi-square test calculation according to the prediction sample and the first prediction data to obtain the significance level value.

[0191] Step S408: When the significance level is greater than the first preset threshold, determine the final output as the first prediction data and the first cumulative distribution function.

[0192] Step S409: Obtain the first real data corresponding to the first prediction data;

[0193] Step S410: Construct a unit step function according to the first cumulative distribution function and the first real data to obtain the first continuous grading probability score.

[0194] Among them, the first continuous ranked probability score (CRPS) can be an index for evaluating the accuracy of probability prediction, used to measure the difference between the prediction distribution and the actual observed values. In the probability prediction of continuous variables, the scoring method is used to evaluate the sharpness of the calibrated prediction distribution. Sharpness refers to the concentration degree of the prediction distribution. A sharp prediction distribution means that the predicted values are concentrated within a smaller interval, indicating a higher certainty. The model with accurate and concentrated predictions is thus rewarded - obtaining a high score; on the contrary, a flat prediction distribution means that the predicted values are scattered within a larger interval, indicating a lower certainty. The model with inaccurate or overly scattered predictions is thus punished - obtaining a low score.

[0195] The calculation formula of the first continuous ranked probability score is:

[0196]

[0197] Among them, F(x) represents the cumulative distribution function of the load prediction; y represents the historical load data, and the unit step function (Heaviside function) is used to determine whether the cumulative distribution function gives the probability greater than or equal to the real data. It can be expressed by the following equivalent expression:

[0198]

[0199] where X and X ′ represent two copies of independent random variables following the same cumulative distribution function F and having the same first moment.

[0200] Step S411, determine the prediction accuracy of the first cumulative distribution function according to the first continuous ranked probability score.

[0201] Among them, the higher the first continuous ranked probability score, the greater the difference between the prediction distribution and the actual observed values (real data), and the lower the prediction accuracy of the first cumulative distribution function; the lower the first continuous ranked probability score, the smaller the difference between the prediction distribution and the actual observed values (real data), and the higher the prediction accuracy of the first cumulative distribution function.

[0202] It can be seen that through the level of the first continuous ranked probability score, the prediction performance of the first cumulative distribution function can be intuitively reflected, providing an important reference basis for model optimization and selection.

[0203] Step S412, when the significance level is not greater than the first preset threshold, input the prediction sample into the second model to obtain the second prediction data and the second probability density function, and determine the second cumulative distribution function according to the second probability density function, and determine the final output as the second prediction data and the second cumulative distribution function.

[0204] Among them, the second model is the KDE model. The second probability density function can be a probability distribution function generated based on the KDE model, which is used to describe the possible distribution of the predicted data. Through the smoothing process of the load data, the second probability density function can more accurately characterize the distribution characteristics of the load demand, providing a reliable reference basis for subsequent analysis.

[0205] In actual operation, after inputting the prediction samples into the KDE model, the model will output the second predicted data and the second probability density function simultaneously. Specifically, the KDE model generates a non-parametric probability density function by applying the Gaussian kernel function to smooth the load data, thereby realizing flexible prediction of the load demand. This process can not only predict the load value but also generate a flexible probability density function, significantly improving the adaptability and prediction accuracy of the model. The second cumulative distribution function can also be a cumulative distribution function obtained by integrating the second probability density function, which is used to describe the probability that the load prediction value is less than or equal to a specific value.

[0206] In a specific embodiment, the process of calculating the second cumulative distribution function can be realized by numerical integration methods, such as using the trapezoidal rule or Simpson's rule to discretely integrate the second probability density function. This process can ensure that the calculation result of the cumulative distribution function has high accuracy and reliability.

[0207] Through the above steps, not only can the probability distribution characteristics of the second predicted data be obtained, but also the distribution range and possibility of the second predicted data can be more intuitively evaluated, providing strong support for market participants to optimize trading strategies.

[0208] Step S413: Obtain the second real data corresponding to the second predicted data.

[0209] Step S414: Construct a unit step function based on the second cumulative distribution function and the second real data to obtain the second continuous grading probability score.

[0210] Step S415: Determine the prediction accuracy of the second cumulative distribution function according to the second continuous grading probability score.

[0211] Among them, the higher the second continuous grading probability score, the lower the prediction accuracy of the second cumulative distribution function, and the lower the second continuous grading probability score, the higher the prediction accuracy of the second cumulative distribution function. The accuracy evaluation of the second predicted data is similar to that of the first predicted data and will not be elaborated here.

[0212] The following Figure 5 will elaborate on the complete steps of the embodiments of the present application in detail:

[0213] Exemplarily, please refer to Figure 5 ,Figure 5 It is a schematic flowchart of another load demand prediction method provided by an embodiment of the present application. As Figure 5 shown, first, smart meter data is input to obtain load demand data of multiple household users. Subsequently, the load data of multiple household users is aggregated, which can be load data aggregation based on a random method or load data aggregation based on portfolio optimization, to obtain multi-time series samples of aggregated load data. Then, time series decomposition is performed. In the training stage, sampling of a training data set and a validation data set can be performed according to the decomposed data, while in the inference stage, sampling of prediction samples can be directly performed. Subsequently, load prediction is performed based on the ARMA-GARCH model, and further Pearson chi-square test calculation is performed based on the load prediction result to obtain a p-value. It is judged whether the p-value is greater than a first preset threshold. When the p-value is greater than the first preset threshold, it is determined that the output is the prediction result of the ARMA-GARCH model, and finally, load prediction performance analysis is performed. When the p-value is not greater than the first preset threshold, load prediction is performed based on the KDE model, and it is determined that the output is the prediction result of the KDE model, and finally, load prediction performance analysis is also performed.

[0214] Consistent with the embodiment shown above, please refer to Figure 6 , Figure 6 It is a block diagram of the functional units of a load demand prediction device provided by an embodiment of the present application. The load demand prediction device is a part of the above load demand prediction module or the load demand prediction module. As Figure 6 shown, the load demand prediction device 60 includes:

[0215] An acquisition unit 601, configured to acquire first load data, where the first load data is obtained by aggregating multiple second load data, and the multiple second load data are respectively historical load demand data based on time series corresponding to multiple users;

[0216] A processing unit 602, configured to decompose the first load data into time series to obtain seasonal data, trend data, and residual data;

[0217] The processing unit 602 is further configured to extract prediction samples based on the seasonal data, trend data, and residual data;

[0218] The processing unit 602 is further configured to input the prediction samples into a first model to obtain first prediction data, where the first model is an autoregressive moving average-generalized autoregressive conditional heteroskedasticity model;

[0219] The processing unit 602 is further configured to perform Pearson chi-square test calculation based on the prediction samples and the first prediction data to obtain a significance level value;

[0220] The processing unit 602 is further configured to determine that the final output is the first prediction data when the significance level is greater than the first preset threshold;

[0221] The processing unit 602 is further configured to, when the significance level is not greater than the first preset threshold, input the prediction sample into the second model to obtain the second prediction data, and determine that the final output is the second prediction data, where the second model is a kernel density estimation model.

[0222] In a feasible embodiment, before obtaining the first load data, the obtaining unit 601 is further configured to:

[0223] Obtain multiple fourth load data;

[0224] Assign a random number to each of the multiple fourth load data, and the multiple random numbers corresponding to the multiple fourth load data are uniformly distributed;

[0225] Determine multiple second load data from the multiple fourth load data, and the random number corresponding to each second load data in the multiple second load data is greater than the second preset threshold.

[0226] In a feasible embodiment, before obtaining the first load data, the obtaining unit 601 is further configured to:

[0227] Obtain multiple fourth load data;

[0228] Divide the first load data range corresponding to the multiple fourth load data into multiple second load data ranges, and obtain the load data combination corresponding to each second load data range in the multiple second load data ranges;

[0229] For each load data combination in the multiple load data combinations, select the first load data combination from each load data combination based on the load data selection vector, where the load data selection vector determines the selection status and non-selection status of each load data in the load data combination based on the first value and the second value respectively, and the first value and the second value are different;

[0230] Perform time series decomposition on the load data in the first load data combination to obtain a seasonal data combination, a trend data combination, and a residual data combination;

[0231] Construct an objective function based on the relative standard deviation between each piece of combined data in the combined data combination and the relative standard deviation between each piece of data in the residual data combination, where the combined data combination is obtained by combining the load data corresponding to the seasonal data combination and the trend data combination;

[0232] Determine the first constraint condition, where the first constraint condition is used to constrain the product of the load data selection vector and the load data vector to be within a preset range;

[0233] Minimize the objective function according to the first constraint condition to obtain the target load data selection vector;

[0234] Select the second load data combination from each load data combination based on the target load data selection vector;

[0235] Determine the load data in the multiple second load data combinations as multiple second load data.

[0236] In a feasible embodiment, in terms of inputting the prediction sample into the first model to obtain the first prediction data, the processing unit 602 is specifically configured to:

[0237] Input the prediction sample into the first model to obtain the first prediction data and the first conditional variance;

[0238] After inputting the prediction sample into the first model to obtain the first prediction data and the first conditional variance, the processing unit 602 is further configured to:

[0239] Generate the first probability density function according to the first prediction data and the first conditional variance;

[0240] Determine the first cumulative distribution function according to the first probability density function.

[0241] In a feasible embodiment, in terms of inputting the prediction sample into the second model to obtain the second prediction data, the processing unit 602 is specifically configured to:

[0242] Input the prediction sample into the second model to obtain the second prediction data and the second probability density function;

[0243] After inputting the prediction sample into the second model to obtain the second prediction data and the second probability density function, the processing unit 602 is further configured to:

[0244] Determine the second cumulative distribution function according to the second probability density function.

[0245] In a feasible embodiment, after determining the first cumulative distribution function according to the first probability density function, the acquisition unit 601 is further configured to: acquire the first real data corresponding to the first prediction data;

[0246] The processing unit 602 is further configured to:

[0247] Construct a unit step function according to the first cumulative distribution function and the first real data to obtain the first continuous ranked probability score;

[0248] Determine the prediction accuracy of the first cumulative distribution function according to the first continuous grading probability score. The higher the first continuous grading probability score, the lower the prediction accuracy of the first cumulative distribution function; the lower the first continuous grading probability score, the higher the prediction accuracy of the first cumulative distribution function.

[0249] In a feasible embodiment, after obtaining the first prediction data, the acquisition unit 601 is further configured to:

[0250] Obtain the first true data corresponding to the first prediction data;

[0251] The processing unit 602 is further configured to:

[0252] Determine the first mean absolute error between the first prediction data and the first true data;

[0253] Determine the accuracy of the first prediction data according to the magnitude of the first mean absolute error. The higher the first mean absolute error, the lower the accuracy of the first prediction data; the lower the first mean absolute error, the higher the accuracy of the first prediction data.

[0254] In a feasible embodiment, in terms of extracting prediction samples based on seasonal data, trend data, and residual data, the processing unit 602 is specifically configured to: perform weighted fusion on the seasonal data, trend data, and residual data to obtain the third load data; extract prediction samples based on the third load data.

[0255] In a feasible embodiment, in terms of extracting prediction samples based on seasonal data, trend data, and residual data, the processing unit 602 is specifically configured to: splice the seasonal data, trend data, and residual data to obtain a three-dimensional vector, analyze the three-dimensional vector based on the self-attention mechanism to obtain three sets of feature data. The first feature data in the three sets of feature data is the feature obtained after determining the weight magnitudes of the query and key based on the seasonal data and the trend data respectively. The second feature data in the three sets of feature data is used to represent the feature obtained after determining the weight magnitudes of the query and key based on the seasonal data and the residual data. The third feature data in the three sets of data is used to represent the feature obtained after determining the weight magnitudes of the query and key based on the residual data and the trend data. Perform weighted fusion on the three sets of feature data to obtain the fourth load data, and extract prediction samples based on the fourth load data.

[0256] It can be understood that since the method embodiment and the device embodiment are different presentation forms of the same technical concept, the content of the method embodiment part in this application should be synchronously adapted to the device embodiment part, and will not be elaborated here.

[0257] In the case of adopting an integrated unit, as Figure 7 shownFigure 7 This is a block diagram of the functional units of another load demand prediction device 60 provided by an embodiment of the present application. In Figure 7 , the load demand prediction device 60 includes: a processing module 712 and a communication module 711. The processing module 712 is used to control and manage the operations of the load demand prediction device 60. For example, steps of an acquisition unit 601 and a processing unit 602, and / or other processes for implementing the technologies described herein. The communication module 711 is used to support the interaction between the load demand prediction device 60 and other devices. As Figure 7 shown, the load demand prediction device 60 may further include a storage module 713, and the storage module 713 is used to store the program code and data of the load demand prediction device 60.

[0258] Among them, the processing module 712 may be a processor or a controller. For example, it may be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logic blocks, modules, and circuits described in connection with the disclosure of the present application. The processor may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and so on. The communication module 711 may be a transceiver, an RF circuit, or a communication interface, etc. The storage module 713 may be a memory.

[0259] Among them, all relevant contents of each scenario involved in the above method embodiment can be cited in the function descriptions of the corresponding functional modules, and will not be elaborated here. The above load demand prediction device 60 can all execute the above Figure 2 shown load demand prediction method.

[0260] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired or wireless manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more collections of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0261] Figure 8 The block diagram of an electronic device provided by an embodiment of the present application. As Figure 8 shown, the electronic device 800 may include one or more of the following components: a processor 801, a memory 802, and a communication interface 803. The processor 801, the memory 802, and the communication interface 803 are interconnected and perform communication with each other. The memory 802 may store one or more computer programs, and the one or more computer programs may be configured to be executed by one or more processors 801 to implement the methods described in the above embodiments.

[0262] The processor 801 may include one or more processing cores. The processor 801 connects various parts within the entire electronic device 800 using various interfaces and lines. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 802, and by invoking data stored in the memory 802, it performs various functions of the electronic device 800 and processes data. Optionally, the processor 801 may be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). The processor 801 may integrate a combination of one or several of a central processing unit (CPU), a graphics processing unit (GPU), and a modem, etc. It can be understood that the above-mentioned modem may not be integrated into the processor 801 and may be implemented separately by a communication chip.

[0263] The memory 802 may include a random access memory (RAM), or may also include a read-only memory (ROM). The memory 802 is used to store instructions, programs, code, code sets, or instruction sets. The memory 802 may include a program storage area and a data storage area. Among them, the program storage area may store instructions for implementing the operating system, instructions for implementing at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc. The data storage area may also store data created during the use of the electronic device 800.

[0264] It can be understood that the electronic device 800 may include more or fewer structural elements than those in the above structural block diagram. For example, it includes a power module, physical buttons, a WiFi (Wireless Fidelity) module, a speaker, a Bluetooth module, sensors, etc., which are not limited herein.

[0265] The above-mentioned electronic device 800 may be a load demand prediction module or a part of a load demand prediction module.

[0266] An embodiment of the present application provides a computer-readable storage medium. Among them, program data is stored in the computer-readable storage medium. When the program data is executed by a processor, it is used to execute some or all of the steps of any one of the load demand prediction methods described in the above method embodiments.

[0267] An embodiment of the present application also provides a computer program product, including a computer program, which is operable to cause a computer to execute some or all of the steps of any one of the load demand prediction methods described in the above method embodiments. The computer program product may be a software installation package.

[0268] It should be noted that, for any of the method embodiments of the above-mentioned load demand prediction methods, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the present application.

[0269] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure content, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of situations. Certain measures are recited in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good effects.

[0270] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of any of the above-mentioned method embodiments of the load demand prediction method can be completed by a program instructing relevant hardware. The program can be stored in a computer-readable memory, and the memory may include: a flash drive, a read-only memory (abbreviation: ROM), a random access memory (abbreviation: RAM), a magnetic disk, or an optical disc, etc.

[0271] The above has introduced the embodiments of the present application in detail. Specific examples are used herein to elaborate on the principles and implementation manners of a load demand prediction method, device, electronic device, and storage medium of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of a load demand prediction method, device, electronic device, and storage medium of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

[0272] This application is described with reference to the flowcharts and / or block diagrams of the methods, hardware products, and computer program products of the embodiments of this application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or a device for implementing the functions specified in one or more of the blocks.

[0273] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or a device for implementing the functions specified in one or more of the blocks.

[0274] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or a device for implementing the functions specified in one or more of the blocks.

[0275] It can be understood that any product that is controlled or configured to execute the processing method of the flowchart described in the method embodiment of a load demand prediction method of this application, such as the terminal of the above flowchart and the computer program product, belongs to the scope of the related products described in this application.

[0276] Obviously, those skilled in the art can make various changes and modifications to a load demand prediction method, device, electronic device, and storage medium provided by this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, this application also intends to include these changes and modifications.

Claims

1. A load demand forecasting method, characterized in that, The method includes: Obtaining first load data, where the first load data is aggregated from a plurality of second load data, and the plurality of second load data are respectively historical load demand data based on time series corresponding to a plurality of users; Performing time series decomposition on the first load data to obtain seasonal data, trend data, and residual data; Extracting prediction samples based on the seasonal data, trend data, and residual data; Inputting the prediction samples into a first model to obtain first prediction data, where the first model is an autoregressive moving average - generalized autoregressive conditional heteroskedasticity model; Performing Pearson chi - square test calculation based on the prediction samples and the first prediction data to obtain a significance level value; When the significance level is greater than a first preset threshold, determining the final output as the first prediction data; When the significance level is not greater than the first preset threshold, inputting the prediction samples into a second model to obtain second prediction data, and determining the final output as the second prediction data, where the second model is a kernel density estimation model.

2. The method according to claim 1, wherein Before obtaining the first load data, the method further includes: Obtaining a plurality of fourth load data; Assigning a random number to each of the plurality of fourth load data, and the plurality of random numbers corresponding to the plurality of fourth load data are uniformly distributed; Determining a plurality of second load data from the plurality of fourth load data, where the random number corresponding to each second load data in the plurality of second load data is greater than a second preset threshold.

3. The method according to claim 1, characterized in that, Before obtaining the first load data, the method further includes: Obtaining a plurality of fourth load data; Dividing the first load data range corresponding to the plurality of fourth load data into a plurality of second load data ranges to obtain a load data combination corresponding to each second load data range in the plurality of second load data ranges; For each load data combination among the plurality of load data combinations, selecting a first load data combination from each load data combination based on a load data selection vector, where the load data selection vector determines the selection status and non - selection status of each load data in the load data combination based on a first value and a second value respectively, and the first value and the second value are different; Performing time series decomposition on the load data in the first load data combination to obtain a seasonal data combination, a trend data combination, and a residual data combination; Constructing an objective function based on the relative standard deviation between each pair of merged data in the merged data combination and the relative standard deviation between each pair of data in the residual data combination, where the merged data combination is obtained by merging the load data corresponding to the seasonal data combination and the trend data combination; Determining a first constraint condition, where the first constraint condition is used to constrain the product of the load data selection vector and the load data vector to be within a preset range; Minimizing the objective function according to the first constraint condition to obtain an objective load data selection vector; Selecting a second load data combination from each load data combination based on the objective load data selection vector; Determine the load data in multiple second load data combinations as the multiple second load data.

4. The method according to any one of claims 1 to 3, characterized in that, The step of inputting the prediction sample into the first model to obtain the first prediction data includes: Input the prediction sample into the first model to obtain the first prediction data and the first conditional variance; After inputting the prediction sample into the first model to obtain the first prediction data and the first conditional variance, the method further includes: Generate a first probability density function according to the first prediction data and the first conditional variance; Determine a first cumulative distribution function according to the first probability density function.

5. The method according to any one of claims 1 to 3, characterized in that, The step of inputting the prediction sample into the second model to obtain the second prediction data includes: Input the prediction sample into the second model to obtain the second prediction data and the second probability density function; After inputting the prediction sample into the second model to obtain the second prediction data and the second probability density function, the method further includes: Determine a second cumulative distribution function according to the second probability density function.

6. The method according to claim 4, wherein After determining the first cumulative distribution function according to the first probability density function, the method further includes: Obtain the first true data corresponding to the first prediction data; Construct a unit step function according to the first cumulative distribution function and the first true data to obtain a first continuous ranked probability score; Determine the prediction accuracy of the first cumulative distribution function according to the first continuous ranked probability score. The higher the first continuous ranked probability score, the lower the prediction accuracy of the first cumulative distribution function, and the lower the first continuous ranked probability score, the higher the prediction accuracy of the first cumulative distribution function.

7. The method according to any one of claims 1 to 3, characterized in that, After obtaining the first prediction data, the method further includes: Obtain the first true data corresponding to the first prediction data; Determine the first mean absolute error between the first prediction data and the first true data; Determine the accuracy of the first prediction data according to the magnitude of the first mean absolute error. The higher the first mean absolute error, the lower the accuracy of the first prediction data, and the lower the first mean absolute error, the higher the accuracy of the first prediction data.

8. A load demand forecasting device, characterized in that, The device includes: An acquisition unit, configured to acquire first load data, where the first load data is aggregated according to multiple second load data, and the multiple second load data are respectively historical load demand data based on time series corresponding to multiple users; A processing unit, configured to perform time series decomposition on the first load data to obtain seasonal data, trend data, and residual data; The processing unit is further configured to extract prediction samples based on the seasonal data, trend data, and residual data; The processing unit is further configured to input the prediction sample into a first model to obtain first prediction data, where the first model is an autoregressive moving average-generalized autoregressive conditional heteroskedasticity model; The processing unit is further configured to perform Pearson chi-square test calculation according to the prediction sample and the first prediction data to obtain a significance level value; The processing unit is further configured to determine that the final output is the first prediction data when the significance level is greater than a first preset threshold; The processing unit is further configured to, when the significance level is not greater than the first preset threshold, input the prediction sample into a second model, obtain second prediction data, and determine that the final output is the second prediction data, where the second model is a kernel density estimation model.

9. An electronic device, the device comprising a processor, a memory, and a computer program stored on the memory, characterized in that, The processor is configured to retrieve the executable program code computer program stored on the memory to execute the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the method according to any one of claims 1-7 is implemented.

Images