Power load prediction method and device and readable storage medium

Abstract

The invention provides a power load prediction method and device and a readable storage medium, and relates to the technical field of power. The method comprises the following steps: acquiring historical data of actual power loads at a plurality of historical moments, performing prediction based on the historical data to obtain a plurality of predicted power loads satisfying normal distribution at a future moment, and determining a final predicted power load at the future moment according to the plurality of predicted power loads. According to the prediction method provided by the invention, single-point prediction can be changed into probability prediction, so that the deviation between the predicted power load and the actual power load at the future moment can be eliminated or reduced.

Description

Technical Field

[0001] The present application relates to the field of electric power technology, and in particular to a method, device and readable storage medium for predicting electric power load. Background Art

[0002] Power load forecasting is the process of predicting the power load of the power supply system at a future time based on the power load of the power supply system at a historical time. The predicted power load is used to control the operation of the power supply system so that the power supply system can provide stable and reliable electricity in the future.

[0003] Generally speaking, power load is affected by many factors such as time and weather, and there is a certain degree of fluctuation. It is usually difficult to accurately capture the fluctuation of power load during the forecasting process, which may cause a large deviation between the predicted power load and the actual power load. Summary of the invention

[0004] The embodiments of the present application provide a method, device and readable storage medium for predicting power load, which can eliminate or reduce the deviation between the predicted power load and the actual power load in the power load prediction process.

[0005] In a first aspect, a method for predicting power load is provided, comprising:

[0006] Acquiring historical data, wherein the historical data includes actual power loads at multiple historical moments;

[0007] Predicting multiple predicted power loads at future times based on the historical data, wherein the multiple predicted power loads satisfy a normal distribution;

[0008] The final predicted power load at the future time is determined according to the multiple predicted power loads.

[0009] In the embodiment of the present application, in the process of predicting the power load, multiple predicted power loads that meet the normal distribution at future moments are predicted based on the actual power load predictions at multiple historical moments, and the final predicted power load at future moments is determined based on the multiple predicted power loads that meet the normal distribution. In this way, the single-point prediction can be changed to a probability prediction to obtain multiple different but related predicted power loads at future moments. The multiple predicted power loads at future moments satisfy the normal distribution and can cover the fluctuation of the power load. Therefore, the final predicted power load at a future moment determined based on multiple predicted power loads has a higher accuracy, which can eliminate or reduce the deviation between the predicted power load and the actual power load at the future moment.

[0010] Furthermore, when the deviation between the predicted power load at a future moment and the actual power load is reduced or eliminated, the power supply system can be controlled more accurately according to the predicted power load, thereby improving the matching degree between the amount of electricity provided by the power supply system and the actual power load, so that the power supply system can provide stable and reliable electricity to electrical equipment, and the electrical equipment can operate continuously and stably.

[0011] In some embodiments, the multiple predicted power loads at future times predicted based on the historical data include: predicting the multiple predicted power loads based on the historical data and first external environmental information, wherein the first external environmental information includes the predicted external environmental information at the future times.

[0012] In an embodiment of the present application, in the process of predicting the power load, the power load at a future time is predicted based on multiple actual power loads and external environmental information at the future time, which can improve the accuracy of the prediction result, thereby eliminating or reducing the deviation between the predicted power load at the future time and the actual power load.

[0013] In some embodiments, the multiple predicted power loads are predicted based on the historical data and the first external environment information, including: inputting the historical data into a variational autoencoder multiple times to obtain the multiple predicted power loads, the decoder in the variational autoencoder is a decoder based on a recurrent neural network, and the first external environment information is the time step input of the decoder in each decoding process.

[0014] In an embodiment of the present application, in the process of predicting the power load, when the first external environmental information is used as the time step input of the decoder in the decoding process, the influence of the output results of the adjacent time steps on the current time step can be reduced, thereby reducing or eliminating the accumulation of deviations, and further eliminating or reducing the deviation between the predicted power load and the actual power load.

[0015] In some embodiments, before predicting multiple predicted power loads at future moments based on the historical data, the method also includes: unifying the orders of magnitude of the actual power loads at the multiple historical moments into a target order of magnitude interval, and the target order of magnitude interval is the order of magnitude interval of the power load used by the variational autoencoder during training.

[0016] In the embodiment of the present application, when using variational autoencoders for power load forecasting, the order of magnitude of multiple actual power loads in historical data can be unified to the target order of magnitude interval during the forecasting process, and then the historical data can be input into the variational autoencoder for forecasting. The order of magnitude of the actual power load input into the variational autoencoder during the training process and the forecasting process is within the same data order of magnitude interval, which can reduce or eliminate the large deviation between the predicted power load and the actual power load at the future moment that may be caused by the inconsistent order of magnitude.

[0017] In some embodiments, unifying the orders of magnitude of the actual power loads at the multiple historical moments into a target order of magnitude range includes: reducing the actual power loads at the multiple historical moments by a target multiple so that the actual power loads at the multiple historical moments are no higher than an upper limit of the target order of magnitude range.

[0018] In an embodiment of the present application, when unifying the orders of magnitude of actual power loads at multiple historical moments, the actual power loads at multiple historical moments are mapped to below the upper limit of the target order of magnitude interval through target multiples, making the process of unifying the orders of magnitude simple and quick, thereby improving the prediction efficiency in the prediction process.

[0019] In some embodiments, before predicting a plurality of predicted power loads at future moments based on the historical data, the method further includes: smoothing the actual power loads at the plurality of historical moments.

[0020] In an embodiment of the present application, multiple actual power loads in the historical data are smoothed during the prediction process, which can eliminate or reduce noise that may exist in the historical data, making the historical data smooth and stable, so that more accurate predictions can be made during the prediction process based on the smoothed and stable historical data, thereby reducing or eliminating the deviation between the predicted power load and the actual power load at a future moment.

[0021] In some embodiments, the smoothing of the actual power loads at the multiple historical moments includes: updating the actual power loads at the other historical moments according to the difference between the actual power loads at the first historical moment and the actual power loads at the other historical moments, wherein the first historical moment is the first of the multiple historical moments.

[0022] In an embodiment of the present application, during the smoothing process, the first actual power load among multiple actual power loads is used as an anchor point, and other actual power loads are updated according to the difference between the first actual power load and the other actual power loads, which can simplify the smoothing process and improve the smoothing efficiency.

[0023] In some embodiments, before predicting a plurality of predicted power loads at future moments based on the historical data, the method further includes: performing denoising processing on the actual power loads at the plurality of historical moments.

[0024] In an embodiment of the present application, during the power load prediction process, historical data is denoised to remove random noise that may exist in the historical data, thereby improving the accuracy of the prediction results, and further reducing or eliminating the possible deviation between the predicted power load and the actual power load.

[0025] In some embodiments, the method of denoising the actual power loads at the multiple historical moments includes one of wavelet denoising, Gaussian filtering, Kalman filtering and mean filtering.

[0026] In some embodiments, the method further includes: determining confidence intervals of the multiple predicted power loads; and determining the reliability of the final predicted power load at the future moment according to the confidence intervals.

[0027] In an embodiment of the present application, the reliability of the final predicted power load at a future moment is determined based on the confidence intervals of multiple predicted power loads, and the prediction results can be evaluated so that certain measures can be taken when the reliability of the prediction results is low, thereby reducing or eliminating the losses caused by the low reliability of the prediction results.

[0028] In some embodiments, the historical data also includes external environment information at the historical moment.

[0029] In the embodiment of the present application, when the historical data includes external environmental information other than the power load, the power load at the future moment can be predicted by combining the external environmental characteristics of the power supply system and the change of the power load during the prediction process, thereby improving the accuracy of the prediction result, and further eliminating or reducing the deviation between the predicted power load and the actual power load.

[0030] In some embodiments, determining the final predicted power load at the future moment according to the multiple predicted power loads includes: determining the final predicted power load at the future moment according to the average of the multiple predicted power loads.

[0031] In an embodiment of the present application, for multiple predicted power loads at future times, the final predicted power load at the future time is determined according to the average of the multiple predicted power loads, so that a more accurate predicted power load at the future time can be determined, thereby improving the accuracy of the predicted power load.

[0032] In some embodiments, the decoder in the variational autoencoder is a decoder built based on a gated recurrent unit.

[0033] In the embodiment of the present application, the decoder constructed based on GRU can alleviate the gradient vanishing and gradient exploding problems that may be caused by the sequence data input into the self-segmenting encoder. At the same time, the decoder constructed based on GRU can better capture the dependency relationship between data, thereby improving the accuracy of the prediction results.

[0034] In some embodiments, the encoder in the variational autoencoder is an encoder based on an attention mechanism.

[0035] In the embodiment of the present application, when the encoder is an encoder based on the attention mechanism, the encoder can effectively process the sequence data, thereby alleviating the possible problems of gradient vanishing and gradient exploding. At the same time, the encoder based on the attention mechanism can perform efficient parallel computing, thereby improving the prediction efficiency. Moreover, the encoder based on the attention mechanism can learn more complex feature representations in historical data, thereby significantly improving the prediction ability of the variational autoencoder, and thus improving the accuracy of the prediction results.

[0036] In a second aspect, a device for predicting power load is provided, comprising:

[0037] An acquisition module, used to acquire historical data, wherein the historical data includes actual power loads at multiple historical moments;

[0038] A prediction module, used for predicting a plurality of predicted power loads at future moments based on the historical data, wherein the plurality of predicted power loads satisfy a normal distribution;

[0039] A determination module is used to determine the final predicted power load at the future moment according to the multiple predicted power loads.

[0040] In some embodiments, the prediction module is specifically used to predict the multiple predicted power loads based on the historical data and first external environment information, and the first external environment information includes the predicted external environment information at the future moment.

[0041] In some embodiments, the prediction module is specifically used to input the historical data into the variational autoencoder multiple times to obtain the multiple predicted power loads, the decoder in the variational autoencoder is a decoder based on a recurrent neural network, and the first external environment information is the time step input of the decoder during each decoding process.

[0042] In some embodiments, the device also includes a unification module for unifying the orders of magnitude of the actual power loads at the multiple historical moments into a target order of magnitude interval, wherein the target order of magnitude interval is the order of magnitude interval of the power load used by the variational autoencoder during training.

[0043] In some embodiments, the unification module is specifically used to reduce the actual power loads at the multiple historical moments by a target multiple so that the actual power loads at the multiple historical moments are no higher than the upper limit of the target order of magnitude range.

[0044] In some embodiments, the device further comprises a smoothing module for smoothing the actual power loads at the multiple historical moments.

[0045] In some embodiments, the smoothing module is specifically used to update the actual power load at other historical moments according to the difference between the actual power load at a first historical moment and the actual power load at other historical moments, where the first historical moment is the first of the multiple historical moments.

[0046] In some embodiments, the device further comprises a denoising module for performing denoising processing on the actual power loads at the multiple historical moments.

[0047] In some embodiments, the method of denoising the actual power loads at the multiple historical moments includes one of wavelet denoising, Gaussian filtering, Kalman filtering and mean filtering.

[0048] In some embodiments, the determination module is further used to determine confidence intervals of the multiple predicted power loads; and determine the reliability of the final predicted power load at the future moment according to the confidence intervals.

[0049] In some embodiments, the historical data also includes external environment information at the historical moment.

[0050] In some embodiments, the determination module is specifically configured to determine the final predicted power load at the future moment according to the average of the multiple predicted power loads.

[0051] In some embodiments, the decoder in the variational autoencoder is a decoder built based on a gated recurrent unit.

[0052] In some embodiments, the encoder in the variational autoencoder is an encoder based on an attention mechanism.

[0053] In a third aspect, a readable storage medium is provided, on which a computer program is stored. When the computer program is run on an electric power load prediction device, the electric power load prediction device executes the electric power load prediction method provided in the first aspect.

[0054] In a fourth aspect, a device for predicting electric load is provided, comprising: a processor; a memory; and a computer program, wherein the computer program is stored in the memory, and when the computer program is executed by the processor, the screening device executes the method for predicting electric load provided in the first aspect.

[0055] In a fifth aspect, a computer program product is provided, comprising: a computer program code, which, when executed on a power load prediction device, enables the power load prediction device to execute the power load prediction method provided in the first aspect.

[0056] In a sixth aspect, a chip is provided, comprising: a processor for calling and running a computer program from a memory, so that an electric power load prediction device equipped with the chip executes the electric power load prediction method provided in the aforementioned first aspect.

[0057] It can be understood that the power load prediction device provided by the second aspect, the readable storage medium provided by the third aspect, the computer program product provided by the fifth aspect, and the chip provided by the sixth aspect are all used to execute the power load prediction method provided by the first aspect. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods provided above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A schematic diagram showing an application scenario of a method for predicting power load provided in an embodiment of the present application.

[0059] Figure 2 A schematic diagram of the principle of a variational autoencoder provided in an embodiment of the present application is shown.

[0060] Figure 3 A schematic diagram of the composition of a decoding layer provided in an embodiment of the present application is shown.

[0061] Figure 4 A schematic diagram of the training process of a variational autoencoder provided in the implementation of the present application is shown.

[0062] Figure 5 A flow chart of a method for predicting power load provided in an embodiment of the present application is shown.

[0063] Figure 6 A flow chart of another method for predicting power load provided in an embodiment of the present application is shown.

[0064] Figure 7 A schematic diagram of the structure of a power load prediction device provided in an embodiment of the present application is shown.

[0065] Figure 8 A structural block diagram of a power load prediction device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0066] The technical solution in the present application will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0067] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0068] The term "comprising" herein indicates the presence of the described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections. The terms "comprising", "including", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized. Below, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, unless otherwise stated, the meaning of "multiple" is two or more.

[0069] The term "and / or" in this article is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0070] The power load, also known as the electricity consumption load, refers to the sum of the electric power obtained by all electrical equipment connected to the power supply system (also known as the power system) from the power supply system at a certain moment.

[0071] In general, the power load is affected by many factors such as time and weather, and has a certain degree of volatility. For example, if the power supply system supplies power to residential areas, most lighting equipment in the residential areas will be turned on at night and turned off during the day, so the power load at night is larger and the power load during the day is smaller. For another example, when the temperature is higher, the use of cooling equipment such as air conditioners and fans in residential areas is greater, and the power load is higher; when the temperature is lower, the use of cooling equipment is less, and the power load is lower. For another example, if the power supply system supplies power to an industrial park, on weekdays, most of the power equipment in the park is in operation, and the power load is higher; on holidays, most of the power equipment in the park is in shutdown, and the power load is lower.

[0072] In addition to normal fluctuations, there may also be abnormal fluctuations in power load. For example, when a transmission line fails in a residential area, most of the electrical equipment in the residential area will be cut off, and the power load at this time will be significantly lower than the normal power load. For another example, when a company in an industrial park suddenly receives a large number of orders, it will start a large number of equipment for production at the same time, which will make the power load significantly higher than the normal power load.

[0073] It should be understood that factors that affect power load fluctuations may include, but are not limited to, the weather, temperature, time, and transmission line failures mentioned in the above examples.

[0074] In order to match the power supply system with the power load as much as possible, during the operation of the power supply system, it is necessary to predict the power load of the power supply system at one or more future moments based on the actual power load of the power supply system at multiple historical moments. Then, the power supply system is controlled to generate electricity and / or store electric energy according to the predicted power load, so that the power supply system can match the power load in the future, thereby providing stable and reliable electric energy to the electrical equipment.

[0075] It should be understood that the historical moment may be a moment that has occurred before the current moment, the future moment may be a moment that is located after the current moment and has not yet occurred, and the current moment is a moment predicted based on multiple actual power loads. The actual power load is the power load actually occurring during the operation of the power supply system, and may be the power load obtained by sampling the power load of the power supply system, and the predicted power load is the predicted power load.

[0076] In the related art, a neural network model is generally used to predict the power load of the power supply system. The neural network model is usually a neural network model such as an auto encoder (AE). In the prediction process, the actual power load of the power supply system at multiple historical moments is first obtained, and then the actual power load at multiple historical moments is input into a pre-trained autoencoder to obtain the predicted power load at one or more future moments output by the autoencoder.

[0077] Among them, the autoencoder adopts single-point prediction, which can only encode the actual power load at multiple historical moments into a single point in the latent space, and then decodes the single point to obtain the predicted power load at the future moment.

[0078] In order to improve the accuracy of the prediction results, during the training process of the autoencoder, the trained autoencoder is usually allowed to encode and decode without information loss. However, when single-point prediction is used, encoding and decoding without information loss may cause serious overfitting of the autoencoder, and overfitting will make it difficult for the autoencoder to accurately capture the fluctuation of the power load, which will reduce the accuracy of the autoencoder. The reduction in accuracy may cause a large deviation between the predicted power load output by the autoencoder and the actual power load at the future moment, especially when there are abnormal fluctuations in the actual power load at multiple historical moments, the predicted power load output by the autoencoder will usually deviate significantly from the actual power load at the future moment.

[0079] Obviously, when the deviation between the predicted power load and the actual power load at a future moment is large, if the power supply system operation is controlled according to the predicted power load, the amount of electricity provided by the power supply system will differ greatly from the power load, which may make the power supply system unable to provide stable and reliable electricity to the electrical equipment, and may even affect the normal operation of the electrical equipment.

[0080] In order to solve the above technical problems, an embodiment of the present application provides a method for predicting electric power load. In the prediction process, multiple predicted electric power loads that meet the normal distribution at future moments are obtained based on the actual electric power load predictions at multiple historical moments, and the final predicted electric power load at the future moment is determined based on the multiple predicted electric power loads that meet the normal distribution. In this way, the single-point prediction can be changed to a probability prediction to obtain multiple different but related predicted electric power loads at future moments. The multiple predicted electric power loads at future moments satisfy the normal distribution and can cover the fluctuation of the electric power load. Therefore, the final predicted electric power load at the future moment determined based on multiple predicted electric power loads has a high accuracy, which can eliminate or reduce the deviation between the predicted electric power load and the actual electric power load at the future moment.

[0081] Furthermore, when the deviation between the predicted power load at a future moment and the actual power load is reduced or eliminated, the power supply system can be controlled more accurately according to the predicted power load, thereby improving the matching degree between the amount of electricity provided by the power supply system and the actual power load, so that the power supply system can provide stable and reliable electricity to electrical equipment, and the electrical equipment can operate continuously and stably.

[0082] The following is a brief introduction to the application scenarios to which the power load prediction method provided in the embodiments of the present application is applicable.

[0083] See also Figure 1 , Figure 1 A schematic diagram showing an application scenario of a method for predicting power load provided in an embodiment of the present application is shown. Figure 1 As shown, the application scenario 10 may include a power supply system 11 and power-consuming equipment 12 .

[0084] Among them, the power supply system 11 may include a prediction device 111, a power generation system 112, an energy storage system 113, an inverter 114 and a transformer 115, etc., but not limited to this. The prediction device 111 can be a computing device such as a host computer, a slave computer, a computer and an industrial computer, but not limited to this. The power generation system 112 may include a photovoltaic power generation system, and may also include a wind power generation system, a hydropower generation system and a thermal power generation system, etc., but not limited to this. The energy storage system 113 may include an electrochemical energy storage system with a battery as an energy storage carrier, and may also include a mechanical energy storage system and a thermal energy storage system, etc., but not limited to this.

[0085] The power generation system 112 and the energy storage system 113 are used to provide a DC voltage to the DC bus 116. The inverter 114 can invert the DC voltage on the DC bus 116 into a first AC voltage, and output the first AC voltage to the first AC bus 117. The voltage level of the first AC voltage is, for example, 380V. The transformer 115 can transform the first AC voltage into a second AC voltage, and output the second AC voltage to the second AC bus 118. The voltage level of the second AC voltage is, for example, 220V.

[0086] The electric device 12 may include various household appliances, such as refrigerators, computers, and lighting equipment, but not limited thereto. The electric device may also include various industrial equipment, such as water pumps, fans, heaters, etc., but not limited thereto. The electric device 12 may be connected to the second AC bus 118 to obtain electric energy from the second AC bus 118. Alternatively, the electric device may also be connected to the first AC bus 117 to obtain electric energy from the first AC bus 117.

[0087] The power load prediction method provided in the embodiment of the present application can be executed by the prediction device 111. The prediction device 111 is connected to the first AC bus 117, and can obtain data such as voltage and current on the first AC bus 117, and determine the actual power load of the power supply system 11 at multiple historical moments based on the obtained voltage and current data.

[0088] After obtaining the actual power loads at multiple historical moments, the prediction device 111 can make predictions based on the actual power loads at multiple historical moments, and predict multiple predicted power loads that satisfy the normal distribution at future moments. Then, the prediction device can determine the final predicted power load at the future moment based on the multiple predicted power loads at the future moment. Finally, the prediction device 111 can control the operation of the power generation system 112 and / or the energy storage system 113 based on the final power load at the future moment, so that the power supply system 11 can output stable and reliable electric energy to the power-consuming equipment 12.

[0089] The above are only illustrative examples, and the power supply system 11 may also include other devices. The prediction device 111 may send the predicted power load to other devices, and the other devices control the operation of the power generation system 112 and / or the energy storage system 113. The specific composition and control method of the power supply system may include but are not limited to the above examples.

[0090] It should be understood that the power load prediction method provided in the embodiment of the present application can be applied to Figure 1 The scenario shown can also be applied to other scenarios where power load forecasting is required, including but not limited to the above examples.

[0091] In an optional implementation, a variational auto encoder (VAE) may be installed in the prediction device 111. In the process of predicting the power load, the prediction device 111 may predict multiple predicted power loads that satisfy the normal distribution at future moments based on the actual power loads at multiple historical moments through the variational auto encoder. It should be understood that the prediction device 111 may also predict the power load at future moments in other ways, and this embodiment does not limit this.

[0092] In order to facilitate understanding of the power load prediction method provided in the embodiment of the present application, the principle of the variational autoencoder used in the prediction process is briefly introduced below with reference to the accompanying drawings.

[0093] See also Figure 2 , Figure 2 FIG. 2 shows a schematic diagram of the principle of a variational autoencoder provided in an embodiment of the present application. The variational autoencoder 20 can be installed and run on Figure 1 In the prediction device 111 shown. Figure 2As shown, the variational autoencoder 20 mainly includes an encoder 21 and a decoder 22. The encoder 21 may include an input layer, a coding layer, a pooling layer, a mean layer and a variance layer, and the decoder 22 may include a decoding layer and an output layer.

[0094] In each process of predicting power load, the prediction device may first input historical data into the input layer. The historical data is the data input into the variational autoencoder during the prediction process. The historical data includes at least the actual power load at multiple historical moments.

[0095] The input layer can receive the historical data input by the prediction device, and map the historical data into a high-dimensional vector (also called an embedding vector) and pass it to the encoding layer for encoding. The encoding layer can encode the high-dimensional vector passed by the input layer to obtain an encoding result to map the historical data to a higher-dimensional feature representation. The pooling layer can pool the encoding result output by the encoding layer to obtain a pooling result to reduce the dimension of the encoding result. The mean layer is usually a fully connected layer, which can map the pooling result to a mean vector in the latent space. The variance layer is usually a fully connected layer, which can map the pooling result to a variance vector in the latent space.

[0096] In each prediction process, the prediction device can obtain a sampling result (usually referred to as a random vector or a random latent vector) by sampling from a standard normal distribution or a normal distribution through a reparameterization trick, and obtain a latent state vector (also referred to as a latent variable) in a latent space by combining the random vector, the mean vector and the variance vector, and then input the latent state vector into the decoder 22.

[0097] The decoding layer in the decoder 22 can decode the latent state vector to obtain a decoding result to map the latent state vector to the space where the historical data is located, and the output layer can convert the decoding result into a predicted power load, that is, the decoder 22 can reconstruct based on the latent state vector to obtain a predicted power load similar to the actual power load.

[0098] Optionally, the encoder is an encoder based on an attention mechanism.

[0099] Exemplarily, the encoder can be an encoder based on a self-attention mechanism, or an encoder based on a multi-head self-attention mechanism. Taking the multi-head self-attention mechanism as an example, the encoder can include multiple encoding layers connected in sequence, each encoding layer mainly includes two sublayers, one sublayer is a multi-head self-attention layer, and the other sublayer is a feedforward neural network layer. In each encoding layer, the input end of the multi-head self-attention layer constitutes the input end of the encoding layer, the output end of the multi-head self-attention layer is connected to the input end of the feedforward neural network, and the output end of the feedforward neural network constitutes the output end of the encoding layer. The encoding layer may also include components such as residual connection and layer normalization, but is not limited to this.

[0100] In one implementation, an encoder identical or similar to the encoder in the Transformer model may be used to construct the encoder in the embodiment of the present application. For example, the encoder in the Transformer model typically includes 5 or 6 encoding layers connected in sequence, each encoding layer including a multi-head self-attention layer and a feedforward neural network layer. In the process of constructing the encoder, an encoder identical or similar to the encoder in the Transformer model may be first obtained, and then a pooling layer, a mean layer, a variance layer, etc. may be added after the encoding layer of the obtained encoder to obtain Figure 2 The encoder 21 is shown.

[0101] It should be noted that the encoder may also be other types of encoders that can encode historical data, and the specific composition and construction method of the encoder may include but are not limited to the above examples.

[0102] In some embodiments, the historical data used in the power load forecasting process is sequence data. When longer sequence data is input into the variational autoencoder, the variational autoencoder may experience gradient vanishing and gradient exploding problems.

[0103] In the embodiment of the present application, when the encoder is an encoder based on the attention mechanism, the encoder can effectively process the sequence data, thereby alleviating the possible problems of gradient vanishing and gradient exploding. At the same time, the encoder based on the attention mechanism can perform efficient parallel computing, thereby improving the prediction efficiency. Moreover, the encoder based on the attention mechanism can learn more complex feature representations in historical data, thereby significantly improving the prediction ability of the variational autoencoder, and thus improving the accuracy of the prediction results.

[0104] Optionally, the decoder may be a decoder based on a recurrent neural network (RNN).

[0105] Exemplarily, a decoding layer in a decoder may include a plurality of neurons connected in sequence, a hidden state vector may be used as a hidden state input of a first neuron among the plurality of neurons, and a hidden state output of each neuron may be used as a hidden state input of an adjacent next neuron.

[0106] See also Figure 3 , Figure 3 FIG. 1 shows a schematic diagram of the composition of a decoding layer provided in an embodiment of the present application. Figure 3 As shown, the decoder can be a decoder built based on a gated recurrent unit (GRU), the neurons in the decoding layer are GRUs, the hidden state input of the first GRU is the hidden state vector, and the hidden state inputs of other GRUs are the hidden state outputs of the adjacent previous GRU.

[0107] When the decoder is an RNN-based decoder, the number of neurons in the decoder is related to the number M of power loads that need to be output during each prediction process. Figure 3 As shown, if the number M of predicted power loads to be output each time is equal to 4, 4 GRUs can be set in the encoding layer. That is, 4 predicted power loads are output each time, each predicted power load corresponds to a future moment, and the 4 future moments are 4 consecutive future moments.

[0108] It should be understood that the decoder may also be a decoder constructed based on other types of RNNs, for example, a decoder based on a long short term memory (LSTM) network, and the specific types of decoders may include but are not limited to the above examples.

[0109] In the embodiment of the present application, the decoder constructed based on GRU can alleviate the gradient vanishing and gradient exploding problems that may be caused by the sequence data input into the self-segmenting encoder. At the same time, the decoder constructed based on GRU can better capture the dependency relationship between data, thereby improving the accuracy of the prediction results.

[0110] In order to facilitate understanding of the power load prediction method provided in the embodiment of the present application, the training process of the variational autoencoder used in the prediction process is briefly introduced below with reference to the accompanying drawings.

[0111] See also Figure 4 , Figure 4 FIG. 1 shows a schematic diagram of a variational autoencoder training process provided in the implementation of the present application. The variational autoencoder training process can be performed by Figure 1 The prediction device 111 shown executes, as Figure 4 As shown, the training process of the variational autoencoder may include steps 41 to 47, and the specific steps in the training process are as follows:

[0112] Step 41: Obtain training set and test set.

[0113] The training set includes a certain number of training samples, which are used to train the variational autoencoder; the test set includes a certain number of test samples, which are used to evaluate whether the trained variational autoencoder meets the requirements.

[0114] When the historical data used in the power load forecasting process is sequence data, the training samples and the test samples are all sequence data, and the sequence lengths of the historical data, training samples, and test samples are the same. For example, the historical data, training samples, and test samples respectively include data of N consecutive historical moments (N is an integer greater than 1), and the data of each historical moment is multidimensional, including at least the actual power load at the historical moment, and may also include information such as the date, temperature, and weather at the historical moment, but not limited thereto. The other information in the data of each historical moment, except the actual power load, describes the external environment of the power supply system at the historical moment, which can be called the external environment information of the historical moment.

[0115] Exemplarily, before training, a power supply system may be monitored for a period of time in advance to obtain data of the power supply system at multiple consecutive historical moments. The data of multiple historical moments constitute sequence data, which may be referred to as training data. Afterwards, the training data may be divided into a training set and a test set. For example, during the monitoring process, information such as the power load, date, temperature, and weather of the power supply system may be obtained once every preset time period (e.g., 10 minutes) to obtain data at a historical moment. After continuous monitoring for a period of time, training data consisting of data at multiple historical moments may be obtained.

[0116] The training data can be expressed as: X = {X1, X2, X3, ..., X t ,…,X T}, T represents the amount of data included in the training data, that is, the number of historical moments. t represents a data in the training data, which corresponds to one of the historical moments, X t It is multidimensional and can be expressed as: X t ∈R d , d represents X t The d dimensions correspond to the actual power load, date, temperature, weather and other information at historical moments.

[0117] In some embodiments, the training data may be divided into a training set and a test set according to one of the historical moments. For example, a data X corresponding to the historical moment s in the training data X may be divided into a training set and a test set. s , and multiple data before the historical moment s are divided into the training set X train ={X1,X2,X3,…,X s}, and divide the multiple data after the historical moment s into the test set X test ={X s+1 ,X s+2 ,X s+3 ,…,X T}, the historical moment s can be set according to the number of training samples and test samples required in the training process.

[0118] It should be understood that the training of the variational autoencoder requires the use of a large number of training samples, and the amount of data included in the training data can be specifically set according to the training requirements of the variational autoencoder. In practical applications, the training set and the test set can also be obtained by other means, and the methods for obtaining the training set and the test set may include but are not limited to the above examples.

[0119] Step 42: Obtain training samples and labeled data from the training set.

[0120] The training samples are used to input the variational autoencoder so that the variational autoencoder outputs a prediction result based on the training samples, and the prediction result includes the predicted power load at one or more moments. The labeled data is used to calculate the error of the variational autoencoder in the prediction process with the prediction result, so as to adjust the model parameters of the variational autoencoder according to the error and train the variational autoencoder.

[0121] In some embodiments, the training set X can be obtained by sliding the window train For example, if the size of the sliding window is W, in the process of selecting training samples and annotated data for the first time, the sliding window can be set starting from the first data in the training set, so that the sliding window selects the 1st to Wth data in the training set, and the first N data in the sliding window constitute a training sample, and the last M data are used as the annotated data corresponding to the training sample.

[0122] Afterwards, the sliding window is moved backward by a preset step size (for example, the preset step size is 1). Each time the sliding window is moved, the first N data in the sliding window are used as training samples, and the last M data are used as corresponding labeled data. By moving the sliding window multiple times in a row, multiple training samples and the labeled data corresponding to each training sample can be obtained.

[0123] Among them, the number N corresponds to the number of historical moments included in the historical data, training samples and test samples, that is, the number of actual power loads input into the variational autoencoder each time is N. M is the number of predicted power loads output by the variational autoencoder each time, and M can be 1 or an integer greater than 1.

[0124] It can be understood that, during the training process, the data of the first N historical moments in the sliding window are input into the variational autoencoder as the prediction basis, and the data of the last M historical moments are used as the actual power load of the power supply system at the M consecutive future moments after the first N historical moments. The prediction device can run the variational autoencoder, so that the variational autoencoder predicts the predicted power load at the last M future moments based on the data of the first N historical moments, and then adjusts the parameters of the variational autoencoder according to the error between the M predicted power loads and the M actual power loads included in the labeled data, so as to train the variational autoencoder.

[0125] In an optional implementation, the prediction device may use the above method to continuously obtain multiple training samples and labeled data corresponding to each training sample from the training set, and then continuously train the variational autoencoder based on the obtained multiple training samples and labeled data.

[0126] In another optional implementation, the prediction device may also train the variational autoencoder once based on the acquired training sample and labeled data after acquiring each training sample and labeled data, and then continue to acquire the next training sample and the corresponding labeled data, and again train the variational autoencoder based on the acquired training sample and labeled data, and so on, to complete multiple training of the variational autoencoder.

[0127] The above are merely illustrative examples, and methods for obtaining training samples and corresponding labeled data from a training set may include but are not limited to the above examples.

[0128] Step 43: pre-process the training samples.

[0129] In some embodiments, during the preprocessing of the training samples, the training samples may first be subjected to denoising to filter out the noise that may exist in the training samples. As mentioned above, the training samples include data from N historical moments. During the denoising process, the actual power load at each historical moment may be first extracted from the data at the N historical moments to obtain a first load sequence consisting of the actual power loads at the N historical moments. The first load sequence may be expressed as Then, the first load sequence may be subjected to wavelet transformation to remove random noise that may exist in the first load sequence.

[0130] It is understandable that the noise in the historical data mainly exists in the power load, so in the denoising process, only the actual power load can be denoised. Of course, if other types of data in the historical data have noise, other data can also be denoised.

[0131] The above are only illustrative examples. In the denoising process, in addition to wavelet denoising, the training samples can also be denoised by denoising methods such as mean filtering, Gaussian filtering and Kalman filtering. The denoising methods may include but are not limited to the above examples.

[0132] Among them, there may be random noise in the actual power load at multiple historical moments in the training samples. During the training process of the variational autoencoder, if there is random noise in the training samples, the variational autoencoder may fit the random noise during the training process, resulting in overfitting of the variational autoencoder, which may reduce the robustness of the variational autoencoder, and further may cause a large deviation between the predicted power load output by the variational autoencoder and the actual power load.

[0133] In an embodiment of the present application, during the training process of the variational autoencoder, the training samples are denoised to remove random noise that may exist in the training samples, thereby improving the robustness of the variational autoencoder, so that during the power load prediction process, the large deviation that may exist between the predicted power load output by the variational autoencoder and the actual power load can be reduced or eliminated.

[0134] In some embodiments, after denoising the training samples, the orders of magnitude of the actual power loads at multiple historical moments included in the training samples can be unified into a preset order of magnitude range to unify the scales of multiple actual power loads in the training samples and eliminate the dimensional error between different actual power loads.

[0135] Exemplarily, the order of magnitude interval may be 0 to 10. After performing wavelet transform denoising on the aforementioned first load sequence, the order of magnitude of each actual power load included in the first load sequence may be mapped to between 1 and 10 to unify the order of magnitudes of multiple actual power loads included in the first load sequence. The order of magnitude interval may also be 0 to 5, or 1 to 8, or 2 to 12, and the order of magnitude interval may include but is not limited to the above examples.

[0136] In some embodiments, the actual power loads at multiple historical moments in the training sample can be reduced by a target multiple so that the actual power loads at multiple historical moments in the training sample are not higher than the upper limit of a preset number interval. Exemplarily, before unifying the order of magnitude of the actual power loads at multiple historical moments, the maximum actual power load can first be determined from the training data, and then the multiple (i.e., the target multiple) by which the maximum actual power load needs to be reduced when the maximum actual power load is reduced to below the upper limit of the preset number interval is determined. Afterwards, the actual power load at each historical moment in the training sample can be reduced by a target multiple so that the order of magnitude of the actual power loads at multiple historical moments in the training sample is unified to within a preset order of magnitude interval.

[0137] For example, when the preset order of magnitude interval is 0 to 10, the upper limit of the preset number interval is 10. Assuming that the maximum actual power load in the training data is 1250, it can be determined that the actual power load 1250 can be reduced to below 10 when 1250 is reduced by 1000 times, so the target multiple can be determined to be 1000. In the process of unifying the order of magnitude of the actual power load in the training sample, each actual power load included in the training sample can be divided by 1000 to map each actual power load to between 0 and 10.

[0138] It should be noted that the specific method of unifying the orders of magnitude of actual power loads at multiple historical moments into a preset order of magnitude range may include but is not limited to the above examples.

[0139] In the related art, the Min-Max Normalization method is usually used to normalize the training samples, test samples and historical data. When this method is used for normalization, the minimum and maximum values ​​of the power load in the training samples may differ greatly from the minimum and maximum values ​​of the power load in the historical data, which may cause a large deviation between the predicted power load output by the variational autoencoder and the actual power load.

[0140] For example, if the maximum value of the actual power load included in the multiple training samples used in the training process is 1000, and the maximum value of the actual power load included in the historical data used in the prediction process is 800. In the prediction process, the output layer in the variational autoencoder will be inversely normalized with the maximum value 1000 in the training samples when outputting the predicted power load. There is a large deviation between the maximum value 1000 and the maximum value 800, so there may be a large deviation between the output predicted power load and the actual power load.

[0141] In an embodiment of the present application, during the training process of the variational autoencoder, the order of magnitude of the power load included in the training sample is unified to a preset order of magnitude interval, which can not only enable the variational autoencoder to converge quickly during the training process, but also improve the stability of the variational autoencoder. Further, during the prediction process, the actual power load in the historical data can be first unified to a preset order of magnitude interval, and then the historical data can be input into the variational autoencoder for prediction. The order of magnitude of the power load input into the variational autoencoder during the training process and the prediction process is in the same interval, which can reduce or eliminate the large deviation that may occur due to the non-uniform order of magnitude.

[0142] In some embodiments, after unifying the order of magnitude of the multiple actual power loads included in the training sample, the multiple actual power loads included in the training sample may be smoothed. As described above, after unifying the order of magnitude of the multiple actual power loads included in the first load sequence, the multiple actual power loads in the first load sequence may be smoothed to obtain a second load sequence, which may be expressed as:

[0143] Exemplarily, during the smoothing process, the actual power load at the first historical moment in the first load sequence (the first historical moment is the first of multiple historical moments) can be used as an anchor point to keep the actual power load at the first historical moment unchanged, and at the same time, the actual power load at other historical moments is updated according to the difference between the actual power load at other historical moments and the actual power load at the first historical moment to obtain a second load sequence.

[0144] For example, for the first load sequence after the unified order of magnitude, the first actual power load in the first load sequence (i.e., the actual power load at the first historical moment) can be directly used as the first power load in the second load sequence, and then the difference between the next adjacent actual power load and the first actual power load is calculated, and the difference is used as the second power load in the second load sequence. Similarly, the difference between other actual power loads and the first actual power load is calculated in sequence, and the difference is used as a power load at the corresponding position in the second load sequence to obtain the second load sequence.

[0145] Alternatively, for actual power loads other than the first actual power load, after determining the difference between the other actual power loads and the first actual power load, the product of the difference and a preset coefficient can be calculated and the product can be used as the power load at the corresponding position in the second load sequence.

[0146] In practical applications, other methods may be used to smooth multiple actual power loads after a unified order of magnitude. For example, multiple actual power loads may be smoothed using methods such as sliding average, exponential smoothing, and spline interpolation. The specific methods of smoothing may include but are not limited to the above examples.

[0147] In an embodiment of the present application, during the preprocessing of training samples, multiple actual power loads in the training samples are smoothed, and the noise that may exist in the training samples can be filtered to reduce the impact of the noise on the training process, thereby improving the generalization ability of the variational autoencoder, and further improving the accuracy of the variational autoencoder in the power load prediction process.

[0148] It should be noted that the preprocessing process of the training samples may include but is not limited to the denoising, unifying the order of magnitude and smoothing processing in the above examples. Similarly, for the labeled data corresponding to the training samples, the labeled data may be preprocessed in the same manner.

[0149] Step 44: obtain a prediction result based on the training sample prediction.

[0150] In this embodiment, after preprocessing the training samples, the prediction device can input the preprocessed training samples into the variational autoencoder, so that the variational autoencoder performs prediction based on the training samples to obtain the prediction results output by the variational autoencoder, and the prediction results include the predicted power loads at M future moments.

[0151] like Figure 2 As shown in FIG. 1 , when the encoder adopts an encoder based on the attention mechanism, the input layer in the encoder 21 can map the data of N historical moments included in the training sample into a high-dimensional vector, and obtain the position code of the data of N historical moments in the training sample, and then add the position code to the high-dimensional vector to obtain an input vector (also called an embedding matrix), and then input the input vector into the encoding layer for encoding, and the encoding result output by the encoding layer can be obtained. The encoding result can be expressed as X enc_output ∈R N×D .

[0152] After obtaining the encoding result, the maximum pooling operation can be performed on the encoding result according to the time dimension through the pooling layer to reduce the dimension of the encoding result and obtain the pooling result. The pooling result can be expressed as:

[0153]

[0154] After obtaining the pooling result, the prediction device passes the pooling result to the mean layer, and the mean layer makes the pooling result and the weight matrix W of the mean layer μ ∈R D×DMultiply to map the pooling result to the mean vector μ∈R in the latent space D Similarly, the pooling result is passed to the variance layer, which makes the pooling result equal to the weight matrix W of the variance layer. σ ∈R D×D Multiply to map the pooling result to the variance vector log(σ 2 ). At the same time, the prediction device can randomly sample from the standard normal distribution to obtain a random vector ε∈R D .

[0155] Then, the prediction device can determine the hidden state vector z through the variance vector, the mean vector and the random vector, and the hidden state vector z=μ+ε×σ. Afterwards, the prediction device can input the hidden state vector into the decoder for decoding.

[0156] like Figure 3 As shown, the decoding layer can be composed of 4 GRUs connected in sequence, each GRU outputs a predicted power load at a future moment, and the hidden state vector can be used as the hidden state input of the first GRU to input the hidden state vector into the decoder.

[0157] In some embodiments, during the training process, the external environment information at each historical moment included in the labeled data can be used as feature information of the decoder during the decoding process. Figure 3 As shown in the figure, the labeled data includes data of four consecutive historical moments, and each historical moment corresponds to a GRU. The external environment information (including date, temperature and weather) of the four historical moments included in the labeled data can be mapped to vector 1, vector 2, vector 3 and vector 4 respectively, and each vector is used as the time step input of a corresponding GRU.

[0158] After the hidden state variables are input into the decoder and the external environment information of each historical moment in the labeled data is mapped into a vector and input into the corresponding GRU, the decoding layer decrypts the hidden state vector based on vector 1, vector 2, vector 3 and vector 4 to obtain the decoding results. The decoding results include y1, y2, y3 and y4. y1, y2, y3 and y4 correspond to the predicted power load at a future moment respectively.

[0159] After obtaining the decoding results, the output layer in the decoder can reversely map the decoding results from the preset order of magnitude interval to the actual data space of the power load to obtain the prediction results. The prediction results include the predicted power loads at four future moments, which can be expressed as

[0160] It should be understood that when the decoder is another type of decoder, the external environment information can also be input into the decoder in other ways, so that the decoder can perform decoding based on the external environment information at a future time during the decoding process.

[0161] Step 45: Adjust the model parameters according to the error between the prediction result and the labeled data.

[0162] For example, the actual power load in the labeled data can be obtained from the pre-processed labeled data to obtain a third load sequence consisting of multiple actual power loads in the labeled data. The third load sequence can be expressed as Y dec ={Y1,Y2,Y3,...,Y M}.

[0163] After getting the prediction results, you can With the third load sequence Y dec Calculate the loss function. The loss function may include the prediction error of the variational autoencoder and the Kullback-Leibler (KL) divergence. The KL divergence is used to measure the difference between the posterior probability distribution q(z|x) of the hidden state vector generated by the training sample and its prior probability distribution (i.e., the standard normal distribution p(z)). The loss function is calculated as follows:

[0164] L=L MSE +λ×L KL (1);

[0165] Among them, L is the function value of the loss function, L MSE is the prediction error of the variational autoencoder, L KL is the KL divergence, λ is the coefficient of the KL divergence, and the calculation method of the KL divergence is as follows:

[0166]

[0167] After calculating the function value L of the loss function, the prediction device can calculate the gradient of the function value L relative to each model parameter in the variational autoencoder, and back-propagate the gradient through the optimizer (such as Adam) to update the model parameters and complete one training of the variational autoencoder.

[0168] Step 46: Whether the end condition is met.

[0169] In some embodiments, the end condition may be that the number of remaining training samples is 0. As described above, in the process of obtaining training samples, multiple training samples and corresponding labeled data may be continuously obtained. During the training process, one training sample is selected from multiple training samples each time, and steps 43 to 46 may be performed in sequence for each training sample. In the process of executing step 46, it is determined whether the number of remaining training samples is 0. If the number of remaining training samples is not 0, return to step 42, obtain the next training sample, and continue to train the variational autoencoder. On the contrary, if the number of remaining training samples is 0, execute step 47.

[0170] Alternatively, the termination condition may be that a new training sample cannot be obtained from the training set. As described above, in the process of obtaining training samples, a training sample and corresponding labeled data may be obtained from the training set each time. After each training sample is obtained, the variational autoencoder is trained based on the obtained training sample. After each training is completed, it is determined whether a new training sample can be obtained from the training set. If a new training sample cannot be obtained, step 47 is executed. On the contrary, if a new training sample can be obtained, the variational autoencoder is trained again based on the obtained training sample.

[0171] The above are merely illustrative examples, and the termination conditions may include but are not limited to the above examples.

[0172] Step 47: Evaluate the variational autoencoder using the test set.

[0173] In this embodiment, after the training of the variational autoencoder is completed using the training set, multiple test samples can be obtained from the test set, and the trained variational autoencoder can be evaluated using the multiple test samples. Exemplarily, during the evaluation process, the prediction device can first obtain multiple test samples from the test set, and obtain the labeled data corresponding to each test sample. The method for obtaining the test sample and the labeled data can refer to the method for obtaining the training sample, which will not be described in detail in this embodiment. After obtaining the test sample, the prediction device can input each test sample into the variational autoencoder in turn, obtain the prediction result output by the variational autoencoder, and respectively calculate the error between the prediction result and the labeled data of each test sample.

[0174] Afterwards, the prediction device can determine the number of test samples whose error between the labeled data and the prediction result is less than a preset threshold from the multiple test samples. When the number of the test samples in the multiple test samples accounts for more than the preset proportion, it can be determined that the accuracy of the variational autoencoder meets the requirements, and the variational autoencoder can be applied in normal power load prediction. On the contrary, it is determined that the accuracy of the variational autoencoder does not meet the requirements, and the variational autoencoder can be retrained.

[0175] The above are merely illustrative examples, and specific methods for evaluating the trained variational autoencoder using a test set may include but are not limited to the above examples.

[0176] It should be noted that, during the training process of the variational autoencoder, the variational autoencoder may also be first trained by other computing devices, and then the trained variational autoencoder may be uploaded to the prediction device. The training process of the variational autoencoder may include but is not limited to the above examples.

[0177] Combination of the above Figure 2 , Figure 3 and Figure 4 The principle and training process of the variational autoencoder used in the power load prediction process are introduced. The power load prediction method provided in the embodiment of the present application is described in detail below with reference to the accompanying drawings.

[0178] See also Figure 5 , Figure 5 FIG. 1 is a flow chart of a method 100 for predicting power load provided in an embodiment of the present application. The execution subject of the method 100 for predicting power load may be Figure 1 The prediction device 111 shown. Figure 5 As shown, the prediction method may include steps 110 to 130, and steps 110 to 130 are specifically as follows:

[0179] Step 110: Obtain historical data.

[0180] In some embodiments, during the prediction of the power load, historical data including actual power loads at multiple historical moments may be first obtained to predict the power load at future moments based on the historical data through a pre-trained variational autoencoder. As mentioned above, the historical data is sequence data and may include data from N historical moments, each of which includes the actual power load at the historical moment, and may also include external environmental information such as the date, temperature, and weather at the historical moment.

[0181] For example, before the prediction device predicts the power load of a power supply system, it can first obtain data from N consecutive historical moments before the current moment to obtain historical data. Figure 1 As shown, the prediction device 111 can continuously detect the voltage and current on the first AC bus 117, and calculate the actual power load of the power supply system 10 based on the voltage and current at a preset time interval, and obtain external environmental information at the same time to obtain data at a historical moment. Before starting the prediction, the prediction device 111 can obtain data from N consecutive historical moments before the current moment to obtain historical data composed of data from N historical moments. The method of obtaining historical data may include but is not limited to the above examples.

[0182] Step 120: Obtain multiple predicted power loads at future times based on historical data prediction, and the multiple predicted power loads satisfy a normal distribution.

[0183] Step 130: Determine the final predicted power load at a future time based on the multiple predicted power loads.

[0184] In each prediction process, the power load at one or more future moments after N consecutive historical moments can be predicted. As mentioned above, if the variational autoencoder predicts and outputs one power load each time, the power load at one future moment after N historical moments is predicted each time; if the variational autoencoder predicts and outputs M power loads each time, the power load at M consecutive future moments after N historical moments is predicted each time.

[0185] In some embodiments, after acquiring the historical data, the prediction device may perform multiple predictions based on the historical data through a variational autoencoder to obtain multiple prediction results. Each prediction result includes a predicted power load at a future time. After multiple predictions are performed to obtain multiple prediction results, the multiple prediction results may include multiple predicted power loads at future times.

[0186] Exemplarily, after acquiring the historical data, the prediction device can continuously predict X times based on the historical data. In each prediction process, the prediction device inputs the historical data into the variational autoencoder, and the encoder in the variational autoencoder encodes the historical data and outputs the mean vector and the variance vector. At the same time, the prediction device can randomly sample from the standard normal distribution to obtain a random vector, and then determine the hidden state vector based on the random vector, the mean vector and the variance vector. Afterwards, the prediction device can input the hidden state vector into the decoder for decoding to obtain the predicted power load at M future moments output by the decoder.

[0187] After X consecutive predictions, X prediction results can be obtained, each of which includes M predicted power loads at future moments, and the first predicted power load among the X prediction results corresponds to the (N+1)th future moment, the second predicted power load corresponds to the (N+2)th future moment, and so on. Each future moment corresponds to X predicted power loads. Afterwards, for each future moment, the final predicted power load at the future moment can be determined based on the X predicted power loads corresponding to the future moment.

[0188] The process of obtaining M predicted power loads based on historical data each time by the variational autoencoder can refer to the training process of the variational autoencoder, which will not be described in detail in this embodiment.

[0189] In the embodiment of the present application, in the process of predicting the power load, firstly, the historical data of the actual power load at multiple historical moments is obtained, and multiple predicted power loads satisfying the normal distribution at future moments are predicted based on the historical data, and then the final predicted power load at the future moment is determined based on the multiple predicted power loads. In this way, the single-point prediction can be changed to a probability prediction, and multiple different but related predicted power loads at future moments can be predicted. The multiple predicted power loads at future moments satisfy the normal distribution and can cover the fluctuation of the power load. Therefore, the final predicted power load at the future moment determined based on the multiple predicted power loads is more accurate, and the deviation between the predicted power load at the future moment and the actual power load can be eliminated or reduced.

[0190] Furthermore, when the deviation between the predicted power load at a future moment and the actual power load is reduced or eliminated, the power supply system can be controlled more accurately according to the predicted power load, thereby improving the matching degree between the amount of electricity provided by the power supply system and the actual power load, so that the power supply system can provide stable and reliable electricity to electrical equipment, and the electrical equipment can operate continuously and stably.

[0191] See also Figure 6 , Figure 6 FIG. 2 is a flow chart of another method 200 for predicting power load provided in an embodiment of the present application. Figure 1 The prediction device 111 shown. Figure 6 As shown, the prediction method may include steps 210 to 250, and steps 210 to 250 are specifically as follows:

[0192] Step 210: Obtain historical data.

[0193] In some embodiments, the historical data also includes external environmental information at historical moments. As described above, the historical data may be sequence data similar to the training samples, including actual power loads at multiple historical moments, and external environmental information including date, temperature, and weather at each historical moment. It should be understood that the external environmental information may include, but is not limited to, date, temperature, and weather.

[0194] In an embodiment of the present application, when the historical data includes external environmental information other than the power load, the power load at future times can be predicted by combining the external environmental characteristics of the power supply system with the changes in the power load during the prediction process, thereby improving the accuracy of the prediction results and eliminating or reducing the deviation between the predicted power load and the actual power load.

[0195] Step 220: pre-process the historical data.

[0196] In some embodiments, during the preprocessing of historical data, the orders of magnitude of actual power loads at multiple historical moments can first be unified into a target order of magnitude interval, where the target order of magnitude interval is the order of magnitude interval of the power load used by the variational autoencoder during the training process.

[0197] Exemplarily, during the training process, if the orders of magnitude of multiple actual power loads in the training samples are mapped to a preset order of magnitude interval of 1 to 10 (i.e., the target order of magnitude interval) during the preprocessing of the training samples, then the orders of magnitude of multiple actual power loads included in the historical data can be mapped to the target order of magnitude interval of 1 to 10.

[0198] For example, the step of unifying the order of magnitude of the actual power loads at multiple historical moments into a target order of magnitude interval may include: reducing the actual power loads at multiple historical moments by a target multiple so that the actual power loads at multiple historical moments are not higher than the upper limit of the target order of magnitude interval. The method of unifying the order of magnitude of multiple actual power loads in historical data can refer to the training sample, and this embodiment will not be described in detail here.

[0199] In the embodiment of the present application, when using variational autoencoders for power load forecasting, the order of magnitude of multiple actual power loads in historical data can be unified to the target order of magnitude interval during the forecasting process, and then the historical data can be input into the variational autoencoder for forecasting. The order of magnitude of the actual power load input into the variational autoencoder during the training process and the forecasting process is within the same data order of magnitude interval, which can reduce or eliminate the large deviation between the predicted power load and the actual power load at the future moment that may be caused by the inconsistent order of magnitude.

[0200] Among them, when unifying the orders of magnitude of actual power loads at multiple historical moments, the actual power loads at multiple historical moments are mapped to below the upper limit of the target order of magnitude interval through the target multiples, making the process of unifying the orders of magnitude simple and fast, thereby improving the prediction efficiency in the prediction process.

[0201] In some embodiments, in the process of pre-processing the historical data, after unifying the orders of magnitude of the actual power loads at multiple historical moments, the actual power loads at multiple historical moments may be smoothed.

[0202] Optionally, the step of smoothing the actual power loads at multiple historical moments may include: updating the actual power loads at other historical moments according to the difference between the actual power load at the first historical moment and the actual power loads at other historical moments, the first historical moment being the first of the multiple historical moments. As mentioned above, the historical data includes the actual power loads at N historical moments, and the first historical moment is the first of the N historical moments.

[0203] The process of smoothing the multiple actual power loads included in the historical data is similar to the process of smoothing the multiple actual power loads included in the training samples, and is not described in detail in this embodiment.

[0204] In an embodiment of the present application, multiple actual power loads in the historical data are smoothed during the prediction process, which can eliminate or reduce noise that may exist in the historical data, making the historical data smooth and stable, so that more accurate predictions can be made based on the smoothed and stable historical data during the prediction process, thereby reducing or eliminating the deviation between the predicted power load and the actual power load at a future moment.

[0205] During the smoothing process, taking the first actual power load among multiple actual power loads as an anchor point and updating other actual power loads according to the difference between the first actual power load and other actual power loads can simplify the smoothing process and improve the smoothing efficiency.

[0206] In some embodiments, in the process of preprocessing historical data, the actual power loads at multiple historical moments can be denoised before the orders of magnitude of the actual power loads at multiple historical moments included in the unified historical data to remove random noise that may exist in the actual power loads at multiple historical moments.

[0207] It should be noted that, in the process of preprocessing historical data, it is also possible not to perform denoising on the actual power loads at multiple historical moments. The method for denoising multiple actual power loads in historical data can refer to the method for denoising multiple actual power loads in training samples, which will not be described in detail in this embodiment.

[0208] In an embodiment of the present application, during the power load prediction process, historical data is denoised to remove random noise that may exist in the historical data, thereby improving the accuracy of the prediction results, and further reducing or eliminating the possible deviation between the predicted power load and the actual power load.

[0209] Step 230: Input the preprocessed historical data into the variational autoencoder multiple times to obtain multiple prediction results.

[0210] Optionally, during the prediction process, multiple predicted power loads at future moments may be predicted based on historical data and first external environment information, and the first external environment information includes the predicted external environment information at future moments.

[0211] In some embodiments, when the decoder in the variational autoencoder is a decoder based on a recurrent neural network, historical data can be input into the variational autoencoder multiple times during the prediction process, and the first external environment information at a future moment can be used as the time step input of the decoder in each decoding process.

[0212] like Figure 3 As shown, the time step input of the decoder is the time step input of each neuron in the decoder. Before prediction, the external environment information of M future moments can be predicted first to obtain the external environment information of each future moment (i.e., the first external environment information). In the prediction process, the external environment information of M future moments can be mapped to vector 1, vector 2, vector 3, and vector 4 in sequence, and then vector 1, vector 2, vector 3, and vector 4 are respectively used as the time step input of each GRU to use the external environment information of each future moment as the feature information in the decoding process.

[0213] Exemplarily, the date of a future moment can be determined by deducing the date of a historical moment, the temperature and weather of a future moment can be obtained through weather forecast prediction, and the method of predicting and obtaining external environmental information can include but is not limited to the above examples.

[0214] Among them, the method of using the external environment information included in the historical data as the feature information in the decoding process is the same as the method of using the external environment information of each historical moment of the labeled data as the feature information in the decoding process during the training process, and the embodiments of the present application will not be repeated here.

[0215] In practical applications, when the time step input of the decoder comes from the inside of the decoder, deviation accumulation may occur, and the deviation may even be amplified, causing a large deviation between the predicted power load at the future moment and the actual power load. For example, if the output y1 of the first GRU is used as the time step input of the second GRU, the output y2 of the second GRU is used as the time step input of the third GRU, and the output y3 of the third GRU is used as the time step input of the fourth GRU, then when the predicted power load at the first future moment (i.e., the predicted power load corresponding to y1) deviates, the deviation will be transmitted to the second GRU and amplified by the second GRU, which may cause a large deviation in the second predicted power load (i.e., the predicted power load corresponding to y2). By analogy, the deviation will be amplified by the third GRU and the fourth GRU, resulting in large deviations in the third and fourth predicted power loads. The deviation of the first predicted power load is gradually accumulated and amplified in each subsequent predicted power load, which may cause a large deviation in each subsequent predicted power load.

[0216] Figure 3When the external environment information of the future moment is input into the decoder, the output of each neuron in the decoder (including y1, y2, y3 and y4) can be made not to be the time step input of other neurons, so that each time step input comes from the outside of the decoder. For example, Figure 3 The vector 1, vector 2, vector 3 and vector 4 shown may be constant terms that have no effect on the decoding result. Alternatively, vector 1, vector 2, vector 3 and vector 4 may be vectors mapped from external environment information at M future moments. It should be understood that the time step input in the decoder may also be other inputs from outside the decoder, which may include but are not limited to constant terms or external environment information.

[0217] In an embodiment of the present application, in the process of predicting the power load, the power load at a future time is predicted based on multiple actual power loads and external environmental information at the future time, which can improve the accuracy of the prediction result, thereby eliminating or reducing the deviation between the predicted power load at the future time and the actual power load.

[0218] At the same time, in the process of power load prediction, when the first external environment information is used as the time step input of the decoder in the decoding process, the influence of the output results of adjacent time steps on the current time step can be reduced, thereby reducing or eliminating the accumulation of deviations, and further eliminating or reducing the deviation between the predicted power load and the actual power load.

[0219] Step 240: Determine multiple predicted power loads at future times from multiple prediction results.

[0220] Step 250: Determine the final predicted power load at a future time according to multiple predicted power loads at future times.

[0221] In this embodiment, the prediction device can obtain a prediction result each time after inputting the preprocessed historical data into the variational autoencoder, and each prediction result includes the predicted power loads of M consecutive future moments. After inputting the historical data into the variational autoencoder multiple times, multiple prediction results can be obtained. Afterwards, the prediction device can determine the predicted power load corresponding to each future moment from the multiple prediction results, thereby obtaining multiple predicted power loads at each future moment. Afterwards, the final predicted power load at the future moment can be determined based on the multiple predicted power loads at each future moment.

[0222] Optionally, the step of determining the final predicted power load at a future time according to a plurality of predicted power loads at a future time may include: determining the final predicted power load at a future time according to an average of the plurality of predicted power loads.

[0223] In some embodiments, for each future moment, after obtaining multiple predicted power loads at the future moment, the prediction device calculates the mean of the multiple predicted power loads, and the mean of the multiple predicted power loads can be directly determined as the final predicted power load at the future moment. Alternatively, after obtaining the mean of multiple predicted power loads at the future moment, the product of the mean and a preset coefficient can be calculated, and the product can be used as the final predicted power load at the future moment. It should be understood that after obtaining the mean of multiple predicted power loads at the future moment, the final predicted power load at the future moment can also be determined by other means based on the mean.

[0224] In an embodiment of the present application, for multiple predicted power loads at future times, the final predicted power load at the future time is determined according to the average of the multiple predicted power loads, so that a more accurate predicted power load at the future time can be determined, thereby improving the accuracy of the predicted power load.

[0225] Optionally, the method may further include:

[0226] Determine confidence intervals for multiple forecasted electric loads;

[0227] The reliability of the final predicted power load at future moments is determined based on the confidence interval.

[0228] In some embodiments, the final predicted power load at the future moment can be evaluated based on the confidence intervals of multiple predicted power loads to determine the reliability of the final predicted power load at the future moment. As mentioned above, after multiple predictions, the multiple predicted power loads at the future moment meet the standard normal distribution. At this time, the means and variances of the multiple predicted power loads can be calculated, and the confidence intervals of the multiple predicted power loads can be determined based on the means and variances of the multiple predicted power loads.

[0229] Among them, the smaller the confidence interval, the higher the reliability of the multiple predicted power loads at the future time, so it can be determined that the reliability of the final predicted power load determined based on the multiple predicted power loads is higher. Conversely, the larger the confidence interval, the lower the reliability of the multiple predicted power loads at the future time, so it can be determined that the reliability of the final predicted power load determined based on the multiple predicted power loads is lower.

[0230] As an optional implementation method, the prediction device can determine that the reliability of the final predicted power load at the future moment is low when the confidence interval is greater than a preset threshold. At this time, the prediction device can output an alarm signal to inform the user that the reliability of the final predicted power load at the future moment outputted by the prediction is low.

[0231] Accordingly, the user can judge the reliability of the prediction result according to the alarm signal and adopt corresponding measures. For example, the user can modify the final predicted power load at the future moment according to the experience value. Alternatively, one can be manually selected from multiple predicted power loads at the future moment as the final predicted power load. Specific measures may include but are not limited to the above examples.

[0232] In an embodiment of the present application, the reliability of the final predicted power load at a future moment is determined based on the confidence intervals of multiple predicted power loads, and the prediction results can be evaluated so that certain measures can be taken when the reliability of the prediction results is low, thereby reducing or eliminating the losses caused by the low reliability of the prediction results.

[0233] Combination of the above Figures 1 to 6 The power load prediction method provided by the embodiment of the present application is described in detail. Figure 7 , Figure 8 The power load prediction device provided in the embodiment of the present application is described. It should be understood that Figure 7 , Figure 8 The control device shown can achieve Figure 4 , Figure 5 or Figure 6 One or more steps in the method flow shown are not described in detail here to avoid repetition.

[0234] Figure 7 FIG. 1 is a schematic diagram showing the structure of a power load prediction device provided by an embodiment of the present application. Figure 7 As shown, the prediction device 70 includes an acquisition module 71 , a prediction module 72 and a determination module 73 .

[0235] An acquisition module 71 is used to acquire historical data, where the historical data includes actual power loads at multiple historical moments;

[0236] A prediction module 72, used to predict multiple predicted power loads at future moments based on historical data, wherein the multiple predicted power loads satisfy a normal distribution;

[0237] The determination module 73 is used to determine the final predicted power load at a future moment according to the multiple predicted power loads.

[0238] In some embodiments, the prediction module is specifically used to predict the multiple predicted power loads based on the historical data and first external environment information, and the first external environment information includes the predicted external environment information at the future moment.

[0239] In some embodiments, the prediction module is specifically used to input the historical data into the variational autoencoder multiple times to obtain the multiple predicted power loads, the decoder in the variational autoencoder is a decoder based on a recurrent neural network, and the first external environment information is the time step input of the decoder during each decoding process.

[0240] In some embodiments, the device further includes: a unification module for unifying the orders of magnitude of the actual power loads at the multiple historical moments into a target order of magnitude interval, wherein the target order of magnitude interval is the order of magnitude interval of the power loads used by the variational autoencoder during training.

[0241] In some embodiments, the unification module is specifically used to reduce the actual power loads at the multiple historical moments by a target multiple so that the actual power loads at the multiple historical moments are no higher than the upper limit of the target order of magnitude range.

[0242] In some embodiments, the device further includes: a smoothing module for smoothing the actual power loads at the multiple historical moments.

[0243] In some embodiments, the smoothing module is specifically used to update the actual power load at other historical moments according to the difference between the actual power load at a first historical moment and the actual power load at other historical moments, where the first historical moment is the first of the multiple historical moments.

[0244] In some embodiments, the device further includes: a denoising module for denoising the actual power loads at the multiple historical moments.

[0245] In some embodiments, the method of denoising the actual power loads at the multiple historical moments includes one of wavelet denoising, Gaussian filtering, Kalman filtering and mean filtering.

[0246] In some embodiments, the determination module is further used to determine confidence intervals of the multiple predicted power loads; and determine the reliability of the final predicted power load at the future moment according to the confidence intervals.

[0247] In some embodiments, the historical data also includes external environment information at the historical moment.

[0248] In some embodiments, the determination module is specifically configured to determine the final predicted power load at the future moment according to the average of the multiple predicted power loads.

[0249] In some embodiments, the decoder in the variational autoencoder is a decoder built based on a gated recurrent unit.

[0250] In some embodiments, the encoder in the variational autoencoder is an encoder based on an attention mechanism.

[0251] Figure 8 FIG. 1 shows a structural block diagram of a power load prediction device provided by an embodiment of the present application. Figure 8 As shown, the power load prediction device 80 includes a processor 81 and a memory 82 , and the above-mentioned components can be connected via one or more buses 84 .

[0252] The power load prediction device 80 further includes a computer program 83, which is stored in the memory 82. When the computer program 83 is executed by the processor 81, the power load prediction device 80 performs the above Figure 5 and Figure 6 All relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding physical device, and will not be repeated here.

[0253] An embodiment of the present application also provides a readable storage medium, which includes a computer program. When the computer program is executed on a computer, the computer executes the method provided by the above method embodiment.

[0254] An embodiment of the present application also provides a computer program product comprising instructions, and when the computer program product is run on a computer, the computer is enabled to execute the method provided by the above method embodiment.

[0255] An embodiment of the present application also provides a chip system, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that a power load prediction device equipped with the chip system executes the method provided in the above method embodiment.

[0256] Among them, the chip system may include an input circuit or interface for sending information or data, and an output circuit or interface for receiving information or data.

[0257] It should be understood that in the embodiments of the present application, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0258] It should also be understood that the memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0259] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0260] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0261] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0262] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0263] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0264] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0265] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting power load, characterized in that: include: Acquiring historical data, wherein the historical data includes actual power loads at multiple historical moments; Predicting multiple predicted power loads at future times based on the historical data, wherein the multiple predicted power loads satisfy a normal distribution; The final predicted power load at the future time is determined according to the multiple predicted power loads.

2. The method according to claim 1, characterized in that The method of obtaining a plurality of predicted power loads at future moments based on the historical data includes: The multiple predicted power loads are predicted based on the historical data and first external environment information, and the first external environment information includes the predicted external environment information at the future moment.

3. The method according to claim 2, characterized in that The step of predicting and obtaining the plurality of predicted power loads based on the historical data and the first external environment information includes: The historical data is input into the variational autoencoder multiple times to obtain the multiple predicted power loads, the decoder in the variational autoencoder is a decoder based on a recurrent neural network, and the first external environment information is the time step input of the decoder in each decoding process.

4. The method according to claim 3, characterized in that Before obtaining a plurality of predicted power loads at future moments based on the historical data, the method further includes: The orders of magnitude of the actual power loads at the multiple historical moments are unified into a target order of magnitude interval, where the target order of magnitude interval is the order of magnitude interval of the power loads used by the variational autoencoder during the training process.

5. The method according to claim 4, characterized in that The step of unifying the orders of magnitude of the actual power loads at the multiple historical moments into a target order of magnitude range includes: The actual power loads at the multiple historical moments are all reduced by a target multiple so that the actual power loads at the multiple historical moments are not higher than the upper limit of the target order of magnitude range.

6. The method according to any one of claims 1 to 5, characterized in that Before obtaining a plurality of predicted power loads at future moments based on the historical data, the method further includes: The actual power loads at the multiple historical moments are smoothed.

7. The method according to claim 6, characterized in that The smoothing of the actual power loads at the multiple historical moments includes: The actual power load at the other historical moments is updated according to the difference between the actual power load at the first historical moment and the actual power load at the other historical moments, wherein the first historical moment is the first one of the multiple historical moments.

8. The method according to any one of claims 1 to 7, characterized in that Before obtaining a plurality of predicted power loads at future moments based on the historical data, the method further includes: The actual power loads at the multiple historical moments are subjected to denoising processing.

9. The method according to claim 8, characterized in that The method of performing denoising processing on the actual power loads at the multiple historical moments includes one of wavelet denoising, Gaussian filtering, Kalman filtering and mean filtering.

10. The method according to any one of claims 1 to 9, characterized in that The method further comprises: determining confidence intervals for the plurality of predicted electric loads; The reliability of the final predicted power load at the future moment is determined according to the confidence interval.

11. The method according to any one of claims 1 to 10, characterized in that: The historical data also includes external environment information at the historical moment.

12. The method according to any one of claims 1 to 11, characterized in that The determining the final predicted power load at the future moment according to the multiple predicted power loads includes: The final predicted power load at the future moment is determined according to the average of the multiple predicted power loads.

13. The method according to any one of claims 3 to 5, characterized in that: The decoder in the variational autoencoder is a decoder constructed based on a gated recurrent unit.

14. The method according to any one of claims 3 to 5, characterized in that: The encoder in the variational autoencoder is an encoder based on the attention mechanism.

15. A device for predicting power load, characterized in that: include: An acquisition module, used to acquire historical data, wherein the historical data includes actual power loads at multiple historical moments; A prediction module, used for predicting a plurality of predicted power loads at future moments based on the historical data, wherein the plurality of predicted power loads satisfy a normal distribution; A determination module is used to determine the final predicted power load at the future moment according to the multiple predicted power loads.

16. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed on the power load prediction device, the power load prediction device executes the method according to any one of claims 1 to 14.

Images