A power load prediction method, device, medium and computer program product

Abstract

The present application relates to the technical field of power load prediction, and specifically provides a power load prediction method, device, medium and computer program product, the method comprising: based on the physical quantity type, separating the historical time series data related to the power load into channels to obtain a plurality of single-channel time series data; generating a multi-scale weight of the corresponding channel and a time series segment embedding under the corresponding scale according to the single-channel time series data; inputting the time series segment embeddings of different scales into a pre-trained hybrid expert prediction model respectively to obtain a plurality of expert prediction results; fusing the plurality of expert prediction results into a single-channel prediction result based on the multi-scale weight; and fusing the single-channel prediction results of the plurality of channels to obtain a power load prediction result. The present application constructs a power load prediction model with channel perception, adaptive scale selection and dynamic expert combination, which can effectively balance the prediction accuracy and the calculation overhead.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power load prediction, in particular to a power load prediction method, device, medium and computer program product. BACKGROUND

[0002] As an important supporting technology for smart grid dispatching, energy management system optimization and power market operation, the core of power load prediction is to accurately model the nonlinear relationship, long-short term dependence and multi-scale change structure in multi-source heterogeneous data. With the increasing proportion of renewable energy such as wind and photovoltaic, the power load presents stronger volatility and uncertainty, and the traditional static prediction method is difficult to meet the actual demand.

[0003] With the development of deep learning, recurrent neural networks (such as LSTM, GRU) and Transformer-based models have been gradually applied to power load prediction tasks. Such models show good performance in modeling temporal dependence, especially in dealing with nonlinearity and long-distance dependence, which is superior to traditional methods. However, existing deep models often use flattening or splicing methods to encode the input when dealing with multi-variable coupling problems, which is difficult to explicitly model the dynamic interaction between different variables (such as temperature, humidity, holiday factors, industrial load, etc.) in the load data. In addition, recurrent networks have natural bottlenecks in gradient propagation and parallel computing capability in long sequence prediction, while Transformer-based models have global perception ability, but their computational complexity is , which limits their efficiency and stability in large-scale load sequence modeling, especially in high-frequency sampling load data prediction tasks.

[0004] The Mixture of Experts (MoE) architecture is a deep learning method with model structure plasticity and strong generalization ability. It dynamically routes the input to different expert modules through the introduction of a gating mechanism, thereby improving the model's representation ability for complex patterns. However, existing MoE methods mostly use static gating strategies, which cannot dynamically allocate experts according to the changing characteristics of the load sequence, resulting in a lack of adaptability of the model in dealing with different load patterns (such as peak-valley fluctuations, sudden abnormalities, periodic variations, etc.).

[0005] More importantly, most current MoE models do not consider the heterogeneous change characteristics of power load at multiple time scales. For example, residential electricity load often has strong daily periodicity, industrial load may exhibit hourly or weekly patterns, and weather factors have a longer time lag on load. Traditional models mostly use fixed window or fixed scale convolution to process input sequences, lacking the ability to flexibly select time scales in different scenarios, which limits the modeling effect of the model on complex load patterns. SUMMARY

[0006] The present application aims to overcome the deficiencies in the prior art, and provides a power load prediction method, device, medium and computer program product, which can efficiently capture the optimal time scale of different power load related historical time series data, and significantly improve the prediction accuracy and generalization ability of the hybrid expert model.

[0007] To achieve the above object, the present application is implemented by using the following technical scheme:

[0008] In a first aspect, the present application provides a power load prediction method, comprising:

[0009] Based on the type of physical quantity, the historical time series data related to the power load is separated by channel to obtain a plurality of single-channel time series data;

[0010] According to the single-channel time series data, a multi-scale weight of the corresponding channel and a time series segment embedding at the corresponding scale are generated;

[0011] The time series segment embeddings of different scales are respectively input into a hybrid expert prediction model pre-trained to obtain a plurality of expert prediction results;

[0012] Based on the multi-scale weight, the plurality of expert prediction results are fused into a single-channel prediction result;

[0013] The single-channel prediction results of multiple channels are fused to obtain a power load prediction result.

[0014] Further, the historical time series data related to the power load is separated by channel to obtain a plurality of single-channel time series data, comprising:

[0015] The historical time series data related to the power load is constructed into a three-dimensional tensor , wherein B represents the batch size, L represents the time series length, and C represents the number of channels;

[0016] The three-dimensional tensor is separated by channel;

[0017] The separated data is converted into one-dimensional convolution form to obtain C single-channel time series data with a size of .

[0018] Further, according to the single-channel time series data, a multi-scale weight of the corresponding channel is generated, comprising:

[0019] The time local features of the single-channel time series data are extracted by using one-dimensional convolution;

[0020] The extracted local temporal features are then subjected to global average pooling to obtain the global features of the single-channel time-series data;

[0021] The global features are converted into multi-scale weights for the corresponding channels using an activation function.

[0022] Furthermore, the global features are converted into multi-scale weights for the corresponding channels using an activation function, including:

[0023] ;

[0024] in, Indicates the corresponding channel Multiscale weights; Indicates channel The number of scales; Indicates channel scale The weight, , ; and These are the parameters for linear transformation; This represents the Gumbel-Softmax activation function.

[0025] Further, based on the single-channel time-series data, the corresponding channel's time-series segment embedding at the corresponding scale is generated, including:

[0026] The single-channel time-series data is divided into a sliding window to obtain a set of time-series segments at each scale.

[0027] The time-series segment set is linearly embedded and average pooled to obtain the time-series segment embedding of the corresponding channel at the corresponding scale.

[0028] Furthermore, it also includes: generating expert weights for the corresponding channels at the corresponding scale based on the embedding of time-series segments.

[0029] Furthermore, based on the embedding of time-series segments, expert weights for the corresponding channels at the corresponding scale are generated, including:

[0030] ;

[0031] in, Indicates the corresponding channel In scale The next Expert weights of each expert sub-model. and These are the gating parameters for the expert sub-model. This indicates the number of expert sub-models. .

[0032] Furthermore, time series segments of different scales are embedded and input into a pre-trained hybrid expert prediction model to obtain multiple expert prediction results, including:

[0033] ;

[0034] in, Indicates channel In scale The following are expert predictions; Indicates the first Expert Sub-model Embedding of time segments The prediction results.

[0035] Furthermore, based on the aforementioned multi-scale weights, multiple expert prediction results are fused into a single-channel prediction result, including:

[0036] ;

[0037] in, Indicates expert prediction results Predicted values ​​after transformation by the prediction head of the hybrid expert prediction model; Indicates channel The number of scales; Indicates channel The single-channel prediction results.

[0038] Secondly, the present invention also provides an electronic device, comprising: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the above-described power load forecasting method.

[0039] Thirdly, the present invention also provides a non-transient machine-readable medium storing computer instructions for causing the computer to execute the above-described power load forecasting method.

[0040] Fourthly, the present invention also provides a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, the above-mentioned power load prediction method is implemented.

[0041] The power load forecasting method provided by this invention deeply integrates a dynamic multi-scale routing mechanism with a hybrid expert network, constructing a hybrid expert forecasting model framework with channel awareness, adaptive scale selection, and dynamic expert combination. This framework can automatically match the optimal modeling path for different input features and effectively balance prediction accuracy and computational cost, making it suitable for various complex power load forecasting scenarios. By introducing the Gumbel-Softmax strategy into the router, the model can dynamically select between multiple time scales, effectively capturing sequence dependencies at different time granularities and overcoming the limitations of traditional fixed-scale methods. Furthermore, a gating mechanism is introduced into the expert module to dynamically allocate expert weights based on input features, enabling on-demand activation and combination of experts. This avoids the insufficient representation problem caused by rigid expert partitioning, significantly improving the model's generalization ability and adaptability. Attached Figure Description

[0042] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating a power load forecasting method provided in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the model framework of the power load forecasting method in this embodiment of the invention;

[0045] Figure 3 This is a schematic diagram of the structure of a power load forecasting device provided in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0047] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0048] like Figure 1 and Figure 2As shown, this embodiment of the invention provides a method for predicting power load. Figure 1 This is a flowchart illustrating the power load forecasting method. This flowchart only shows the logical sequence of the method described in this embodiment. In other possible embodiments of the invention, different methods may be used, provided there are no conflicts. Figure 1 Complete the steps shown or described in the order indicated. Figure 2 This is a schematic diagram of the model framework for the power load forecasting method.

[0049] The power load forecasting method provided in this embodiment can be applied to a terminal and can be executed by a power load forecasting device. This device can be implemented in software and / or hardware and can be integrated into the terminal.

[0050] See Figure 1 and Figure 2 The method of this invention specifically includes the following steps:

[0051] Step S101: Based on the physical quantity type, perform channel separation on the historical time series data related to power load to obtain multiple single-channel time series data.

[0052] Historical time-series data related to power load can include historical power load data, as well as corresponding historical meteorological data and historical time data. Specifically, historical power load data generally refers to the actual load values ​​of the area to be predicted and / or its surrounding areas (such as the entire province or a single power plant or wind turbine group) at different historical moments; historical meteorological data refers to environmental variables highly correlated with power load, such as temperature, humidity, wind speed, wind direction, and solar radiation; historical time data includes calendar information (year, month, day, hour), weekday / holiday labels, and seasonal indicators.

[0053] The historical time-series data related to the power load input model is multivariate time-series data, containing variables of multiple different physical quantity types. In this invention, each physical quantity is treated as an independent channel. For example, when the data contains four physical quantities: "power load, temperature, humidity, and wind speed", the number of channels is 4.

[0054] Specifically, historical time-series data related to power load are separated into multiple single-channel time-series data, including:

[0055] Historical time-series data related to power load are constructed into a three-dimensional tensor. Where B represents the batch size, L represents the time series length, and C represents the number of channels;

[0056] For three-dimensional tensors Perform channel separation;

[0057] The separated data is converted into a one-dimensional convolutional form for subsequent convolution and fragment embedding operations, resulting in C data points of size C. Single-channel timing data.

[0058] Step S102: Based on the single-channel time series data, generate the multi-scale weights of the corresponding channels and the embedding of time series segments at the corresponding scales.

[0059] Specifically, based on the single-channel time-series data, multi-scale weights for the corresponding channels are generated, including:

[0060] One-dimensional convolution is used to extract the temporal local features of the single-channel time-series data;

[0061] The extracted local temporal features are then subjected to global average pooling to obtain the global features of the single-channel time-series data;

[0062] The global features are converted into multi-scale weights for the corresponding channels using an activation function.

[0063] In this embodiment of the invention, for each sample in the batch input (i.e., an instance consisting of a historical time series of length L and C channel variables), weights of multiple scales are dynamically generated.

[0064] Single-channel timing data of channel c (c=1,...,C) For example, the specific steps for generating its multi-scale weights include:

[0065] Extracting single-channel time-series data using 1D convolution The time-local information is calculated using the following formula:

[0066] ;

[0067] in, Indicates channel Local features; This represents a one-dimensional convolution operation; This represents the intermediate feature dimension of the convolution output;

[0068] right Perform global average pooling to obtain channels global features The calculation formula is:

[0069] ;

[0070] in, This indicates adaptive global average pooling;

[0071] Multiple scale weights are generated through linear transformation and the Gumbel-Softmax activation function:

[0072] ;

[0073] in, Indicates channel Multiscale weights, Indicates channel The number of scales; Indicates channel scale Scale weights, , ; and These are the parameters for linear transformation; This represents the Gumbel-Softmax activation function.

[0074] The same method can be used to generate multi-scale weights for other channels, which will not be elaborated here.

[0075] On the other hand, based on the single-channel time-series data, the corresponding channel's time-series segment embedding at the corresponding scale is generated, including:

[0076] The single-channel time-series data is divided into a sliding window to obtain a set of time-series segments at each scale.

[0077] The time-series segment set is linearly embedded and average pooled to obtain the time-series segment embedding of the corresponding channel at the corresponding scale.

[0078] For example, generating the above channel c at scale The specific steps for embedding time-series segments include:

[0079] For single-channel timing data Perform sliding window segmentation to obtain channels In scale The following time series fragment set The calculation formula is:

[0080] ;

[0081] in, This indicates a sliding window operation; This represents the window length corresponding to scale i. The possible values ​​are {4, 8, 16, 32}, corresponding to differences in time granularity; .

[0082] right Linear embedding is performed, projecting each time series segment onto a fixed dimension through a linear layer. After that, Perform average pooling to obtain channels In scale Timing segment embedding The formula is expressed as:

[0083] ;

[0084] ;

[0085] in, Represents a linear mapping. This represents the embedding space dimension uniformly set in the prediction framework.

[0086] The same method can be used to generate the temporal segment embeddings of channel c at other scales, which will not be elaborated here.

[0087] Step S103: Embed time series segments of different scales and input them into the pre-trained hybrid expert prediction model to obtain multiple expert prediction results.

[0088] In this embodiment of the invention, the network structure of the expert submodule in the hybrid expert prediction model includes a set of independently trained multilayer perceptrons.

[0089] With the above channels In scale Taking the embedding of time segments as an example, in this embodiment of the invention, step S104 specifically includes:

[0090] According to the channel In scale The following time-series segments are embedded to generate corresponding expert weights, including:

[0091] ;

[0092] in, Indicates channel In scale The next Expert weights of each expert sub-model. and These are the gating parameters for the expert sub-model. Indicates the number of sub-models. .

[0093] In an embodiment of the present invention, The value can be 4 or 8.

[0094] Channel In scale The embedded time series segments are input into a pre-trained hybrid expert prediction model to obtain expert prediction results, calculated using the following formula:

[0095] ;

[0096] in, Indicates channel In scale The following are expert predictions; Indicates the first An expert sub-model embeds time-series segments. The prediction results.

[0097] The same method can be used to obtain the expert prediction results for channel c at other scales, which will not be elaborated here.

[0098] Step S104: Based on the multi-scale weights, merge the prediction results of multiple experts into a single-channel prediction result.

[0099] With the above channels For example, in this embodiment of the invention, step S105 specifically includes:

[0100] Channel The expert prediction results across all scales are weighted and fused according to their respective scale weights, calculated using the following formula:

[0101] ;

[0102] in, Indicates expert prediction results Predicted values ​​after transformation by the prediction head of the hybrid expert prediction model; Indicates channel The number of scales; Indicates channel The single-channel prediction results.

[0103] Step S105: Merge the single-channel prediction results from multiple channels to obtain the power load prediction result.

[0104] In this embodiment of the invention, the prediction results of each single channel are spliced ​​together to obtain the power load prediction result. , .

[0105] The power load forecast result is the power load value for the future forecast period, which can be a single-point forecast or a multi-step forecast.

[0106] The training data for the hybrid expert prediction model of this invention consists of real historical power load data, historical meteorological data, and historical time data. The training objective is to minimize the error between the predicted value and the actual load value. This embodiment of the invention uses the MSE loss function for end-to-end training and the Adam optimizer for model parameter updates. The MSE loss function is:

[0107] ;

[0108] in, Indicates the predicted value. Represents the actual value.

[0109] To verify the effectiveness of the power load forecasting method of the present invention, relevant experiments were conducted.

[0110] Experimental conditions: Model training and evaluation were performed on an NVIDIA GeForce RTX 3090 GPU using the PyTorch deep learning framework. The historical observation window length was set to 96 (corresponding to 1 day of data in the ECL dataset), and prediction step sizes were set to {96, 192, 336, 720}, covering different prediction periods from short-term to long-term. The model used a batch size of 32, the optimizer was Adam, and the initial learning rate was set to 1×10⁻⁶. -4 The learning rate is dynamically adjusted using cosine annealing to improve convergence stability. The maximum training period is 30 epochs, and an early stopping mechanism is introduced during training: if the validation set error does not decrease significantly within 5 consecutive epochs, training is terminated early.

[0111] This experiment used the ECL (Electricity Consumption Load) dataset as the main test object, covering electricity load variation data from approximately 300 users, exhibiting significant periodicity and multivariate time-series dependence characteristics. To comprehensively evaluate the predictive capability of the model in this invention, it was compared with the following mainstream benchmark models: PatchTST, DLinear, FEDformer, Autoformer, Informer, and Pyraformer. Evaluation metrics included mean squared error (MSE) and mean absolute error (MAE); the smaller the values ​​of both, the better the predictive performance.

[0112] As shown in Table 1, under four prediction step sizes (96, 192, 336, and 720) on the ECL dataset, the model of this invention achieves the best or near-best performance in both MSE and MAE, demonstrating excellent prediction accuracy and stability. The model's MSE is lower than other comparative methods across all four prediction periods, and it also leads in almost all MAE metrics, especially exhibiting stronger generalization ability and stability under long-period (336 and 720 step sizes) predictions. This indicates that this method has significant advantages in multi-scale modeling and dynamic expert selection mechanisms for complex power load time series.

[0113] Table 1 Comparison of Model Prediction Performance

[0114]

[0115] The power load forecasting method provided by this invention deeply integrates a dynamic multi-scale routing mechanism with a hybrid expert model, constructing an end-to-end time series modeling framework with channel awareness, adaptive scale selection, and dynamic expert combination. This framework can automatically match the optimal modeling path for different input features. By introducing the Gumbel-Softmax strategy into the router, the model can dynamically select between multiple time scales, effectively capturing sequence dependencies at different time granularities and overcoming the limitations of traditional fixed-scale methods. A gating mechanism is introduced into the expert module, dynamically allocating expert weights based on input features, enabling on-demand activation and combination of experts. This avoids the insufficient representation problem caused by rigid expert partitioning, significantly improving the model's generalization ability and adaptability. While improving the ability to model multivariate time series structures, this invention reduces redundant computation through dynamic routing and expert selection mechanisms, effectively balancing prediction accuracy and computational cost.

[0116] Furthermore, this method has good scalability and versatility, and is applicable not only to power load forecasting, but also to various complex application scenarios such as meteorological modeling and financial analysis. It provides an efficient, flexible and scalable new solution to the problems of fixed-scale modeling, static expert network, and insufficient sequence representation ability in multivariate sequence forecasting tasks.

[0117] Based on the power load forecasting method provided by the embodiments of the present invention, the embodiments of the present invention also provide a power load forecasting device. The solution provided by the device is similar to the solution described in the above method. Therefore, the specific limitations in the embodiments of the power load forecasting device provided below can be found in the limitations of the power load forecasting method above, and will not be repeated here.

[0118] like Figure 3 As shown, an embodiment of the present invention provides a power load forecasting device, comprising:

[0119] The channel separation module is used to separate historical time-series data related to power load based on the type of physical quantity, and obtain multiple single-channel time-series data.

[0120] The multi-scale weight and time series segment embedding acquisition module is used to generate the multi-scale weight of the corresponding channel and the time series segment embedding at the corresponding scale based on the single-channel time series data.

[0121] The expert prediction result acquisition module is used to embed time series segments of different scales into a pre-trained hybrid expert prediction model to obtain multiple expert prediction results.

[0122] The single-channel prediction result acquisition module is used to fuse multiple expert prediction results into a single-channel prediction result based on the multi-scale weights.

[0123] The power load forecasting result acquisition module is used to fuse the single-channel forecasting results from multiple channels to obtain the power load forecasting result.

[0124] An embodiment of the present invention also provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the power load forecasting method of an embodiment of the present invention.

[0125] Embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the power load forecasting method of embodiments of the present invention.

[0126] Embodiments of the present invention also provide an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the power load forecasting method of the embodiments of the present invention.

[0127] The present invention will now describe a structural block diagram of an electronic device that can serve as an embodiment of the present invention, serving as an example of a hardware device applicable to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0128] like Figure 4 As shown, the electronic device includes a computing unit 201, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 202 or a computer program loaded from a storage unit 208 into a random access memory (RAM) 203. The RAM 203 may also store various programs and data required for the operation of the electronic device. The computing unit 201, ROM 202, and RAM 203 are interconnected via a bus 204. An input / output (I / O) interface 205 is also connected to the bus 204.

[0129] Multiple components in the electronic device are connected to I / O interface 205, including: input unit 206, output unit 207, storage unit 208, and communication unit 209. Input unit 206 can be any type of device capable of inputting information into the electronic device. Input unit 206 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 207 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 208 may include, but is not limited to, disks and optical discs. Communication unit 209 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0130] The computing unit 201 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 201 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 201 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 208. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 202 and / or communication unit 209. In some embodiments, the computing unit 201 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0131] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0132] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0133] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0134] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0135] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0136] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0137] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for predicting electricity load, characterized in that, include: Based on the type of physical quantity, the historical time series data related to power load are separated into multiple single-channel time series data. Based on the single-channel time series data, generate multi-scale weights for the corresponding channels and embedding of time series segments at the corresponding scales; Embedding time series segments of different scales and inputting them into a pre-trained hybrid expert prediction model yields multiple expert prediction results. Based on the multi-scale weights, multiple expert prediction results are fused into a single-channel prediction result; The power load forecast results are obtained by fusing the single-channel forecast results from multiple channels. Specifically, based on the single-channel time-series data, multi-scale weights for the corresponding channels are generated, including: One-dimensional convolution is used to extract the temporal local features of the single-channel time-series data; The extracted local temporal features are then subjected to global average pooling to obtain the global features of the single-channel time-series data; The global features are converted into multi-scale weights for the corresponding channels using an activation function, including: ； in, Indicates the corresponding channel Multiscale weights; Indicates channel The number of scales; Indicates channel scale The weight, , ; and These are the parameters for linear transformation; This represents the Gumbel-Softmax activation function; Indicates channel Global features.

2. The power load forecasting method according to claim 1, characterized in that, Historical time-series data related to power load are separated into multiple single-channel time-series data, including: Historical time-series data related to power load are constructed into a three-dimensional tensor. Where B represents the batch size, L represents the time series length, and C represents the number of channels; For three-dimensional tensors Perform channel separation; The separated data is converted into a one-dimensional convolutional form, resulting in C data points of size C. Single-channel timing data.

3. The power load forecasting method according to claim 1, characterized in that, Based on the single-channel time-series data, generate the corresponding channel's time-series segment embedding at the corresponding scale, including: The single-channel time-series data is divided into a sliding window to obtain a set of time-series segments at each scale. The time-series segment set is linearly embedded and average pooled to obtain the time-series segment embedding of the corresponding channel at the corresponding scale.

4. The power load forecasting method according to claim 1, characterized in that, Also includes: Based on the embedding of time-series segments, expert weights for the corresponding channels at the corresponding scale are generated.

5. The power load forecasting method according to claim 4, characterized in that, Based on the embedding of time-series segments, expert weights for the corresponding channels at the corresponding scale are generated, including: ； in, Indicates the corresponding channel In scale The next Expert weights of each expert sub-model. and These are the gating parameters for the expert sub-model. This indicates the number of expert sub-models. ; This indicates the embedding of time sequence segments.

6. The power load forecasting method according to claim 5, characterized in that, Time series segments of different scales are embedded and fed into a pre-trained hybrid expert prediction model to obtain multiple expert prediction results, including: ； in, Indicates channel In scale The following are expert predictions; Indicates the first Expert Sub-model Embedding of time segments The prediction results.

7. The power load forecasting method according to claim 6, characterized in that, Based on the aforementioned multi-scale weights, multiple expert prediction results are fused into a single-channel prediction result, including: ； in, Indicates expert prediction results Predicted values ​​after transformation by the prediction head of the hybrid expert prediction model; Indicates channel The number of scales; Indicates channel The single-channel prediction results.

8. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the power load forecasting method according to any one of claims 1 to 7.

9. A non-transitory machine-readable medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the power load forecasting method according to any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the power load forecasting method according to any one of claims 1 to 7.

Images