Power load and electricity price combined prediction method and system based on multi-task learning

Abstract

The invention relates to the technical field of joint prediction, in particular to a power load and electricity price joint prediction method and system based on multi-task learning. The method comprises the following steps: respectively executing time sequence decomposition on a historical power load sequence and a historical electricity price sequence, constructing a multi-scale joint input feature vector sequence, obtaining shared feature representation, and generating shared features of task deviation by utilizing a learnable channel attention module, and generating a load task characteristic and an electricity price task characteristic which are deeply fused by using a cross-task interaction gating unit, generating a power load and electricity price predicted value, calculating a load predicted loss, an electricity price predicted loss and a relation consistency loss, obtaining a combined total loss function, and updating network parameters. According to the scheme, task conflicts and negative migration can be relieved, the deep relationship between the power load and the electricity price is mined, the consistency of the prediction result and the real situation is ensured, and the difficulty of network learning is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of joint prediction. More particularly, the present application relates to a power load and electricity price joint prediction method and system based on multi-task learning. BACKGROUND

[0002] Power load can reflect the power demand of users, and electricity price is affected by various factors such as supply and demand relationship, power generation cost, market strategy, etc. Power load and electricity price will show mutual influence relationship in time series.

[0003] When researching on the problem of power load and electricity price prediction, researchers usually regard power load and electricity price prediction as two independent single-task prediction problems, i.e. constructing prediction models for the two tasks respectively. However, the above processing method breaks the internal relationship between power load and electricity price, and fails to fully utilize the interaction mechanism of the two in time series.

[0004] In order to utilize the correlation between power load and electricity price, the prior art adopts a multi-task learning framework for joint prediction of power load and electricity price. In the specific network structure, a hard parameter sharing bottom network structure is usually adopted, i.e. two tasks share a feature extractor, and then connect independent prediction heads respectively; this method attempts to realize information sharing between different tasks to a certain extent by sharing the bottom feature extraction network, thereby improving the prediction accuracy.

[0005] However, the method of the multi-task learning framework can link power load and electricity price through information sharing, but still has obvious deficiencies. First, the hard parameter sharing structure forces all tasks to use the same feature representation, which is difficult to distinguish the preference difference of different tasks for shared features, and is prone to cause gradient conflict and negative transfer phenomenon between tasks, thereby affecting the training effect of the joint prediction model. Second, the existing joint prediction model usually constructs a joint loss by simply adding each task loss function, which lacks modeling and constraint of the correlation structure between power load and electricity price prediction sequences, and cannot guarantee that the prediction results are consistent with the real situation in the relationship level. In addition, directly taking the original time series containing multiple complex patterns such as trend, periodicity and random fluctuation as the input of the joint prediction model increases the difficulty of network learning and distinguishing different scale features, and limits the extraction ability of the joint prediction model for complex time series features. SUMMARY

[0006] The purpose of the present application is to propose a power load and price joint prediction method and system based on multi-task learning, to solve the problems that the existing technology is prone to cause gradient conflict and negative transfer phenomenon between different tasks, cannot guarantee that the prediction result is consistent with the real situation in the relationship level, and the joint prediction model constructed has weak feature extraction ability for complex characteristics; therefore, the present application provides a scheme in the following two aspects.

[0007] In the first aspect, the present application provides a power load and price joint prediction method based on multi-task learning, comprising: obtaining a historical power load sequence, a historical price sequence, meteorological data and calendar data; performing time series decomposition on the historical power load sequence and the historical price sequence respectively to obtain their respective trend components, periodic components and residual components; concatenating the decomposed components with the meteorological data and the calendar data to construct a multi-scale joint input feature vector sequence; inputting the multi-scale joint input feature vector sequence into a shared feature extraction network to obtain a shared feature representation; adjusting the channel weight of the shared feature representation by using a learnable channel attention module to generate a task-biased shared feature; inputting the task-biased shared feature into a power load task exclusive network and a price task exclusive network respectively to obtain preliminary power load features and preliminary price features; using a cross-task interaction gate unit to fuse the preliminary power load features and the preliminary price features to generate deeply fused load task features and price task features; based on the deeply fused load task features and the price task features, generating power load prediction values and price prediction values at a target time through a prediction head network; calculating load prediction loss, price prediction loss and relationship consistency loss, and weighting and summing the three to obtain a joint total loss function; updating network parameters based on the joint total loss function.

[0008] Preferably, the time series decomposition on the historical power load sequence and the historical price sequence respectively to obtain their respective trend components, periodic components and residual components comprises: adopting a decomposition algorithm based on local weighted regression scatter smoothing method; decomposing the historical power load sequence input into a load trend component, a load periodic component and a load residual component; decomposing the historical price sequence input into a price trend component, a price periodic component and a price residual component.

[0009] Preferably, the inputting the multi-scale joint input feature vector sequence into a shared feature extraction network to obtain a shared feature representation comprises: the shared feature extraction network adopts a bidirectional gated recurrent unit network; inputting the multi-scale joint input feature vector sequence into the bidirectional gated recurrent unit network to extract time series features from the positive and negative directions through the forward and backward gated recurrent unit layers respectively, and concatenating the forward and backward hidden layer states to form a shared feature representation.

[0010] Preferably, the learning channel attention module adjusts the channel weight of the shared feature representation to generate task-biased shared features, comprising: performing a squeeze operation on the shared feature representation through a global average pooling layer to obtain a channel representation; inputting the channel representation into a bottleneck structure composed of two fully connected layers for excitation operation to learn the weight coefficient of each channel; multiplying the learned weight coefficient with the shared feature representation by channel to generate task-biased shared features.

[0011] Preferably, the task-biased shared features are input into the power load task-specific network and the electricity price task-specific network respectively to obtain preliminary power load features and preliminary electricity price features, comprising: inputting the task-biased shared features into the first fully connected layer as the power load task-specific network and the second fully connected layer as the electricity price task-specific network respectively to obtain preliminary power load features and preliminary electricity price features.

[0012] Preferably, the cross-task interaction gating unit is used to fuse the preliminary power load features and the preliminary electricity price features to generate deep fused load task features and deep fused electricity price features, comprising: generating a load gating signal by passing the preliminary electricity price features through a third fully connected layer and a Sigmoid activation function, and performing element-level multiplication of the load gating signal and the preliminary power load features to obtain load features modulated by electricity price information; generating a electricity price gating signal by passing the preliminary power load features through a fourth fully connected layer and a Sigmoid activation function, and performing element-level multiplication of the electricity price gating signal and the preliminary electricity price features to obtain electricity price features modulated by load information; adding the preliminary power load features and the load features modulated by electricity price information to generate deep fused load task features; adding the preliminary electricity price features and the electricity price features modulated by load information to generate deep fused electricity price features.

[0013] Preferably, based on the deep fused load task features and the deep fused electricity price features, the prediction head network generates power load prediction values and electricity price prediction values at the target time, comprising: the prediction head network is a first multi-layer perceptron and a second multi-layer perceptron with the same structure, and the multi-layer perceptron is composed of two fully connected layers and a linear output layer; inputting the deep fused load task features into the first multi-layer perceptron to output power load prediction values at the target time; inputting the deep fused electricity price features into the second multi-layer perceptron to output electricity price prediction values at the target time.

[0014] Preferably, the method for calculating the relationship consistency loss is as follows: calculate the Pearson correlation coefficient between the predicted power load sequence and the predicted electricity price sequence; calculate the Pearson correlation coefficient between the actual power load sequence and the actual electricity price sequence; and take the absolute value of the difference between the two Pearson correlation coefficients as the relationship consistency loss.

[0015] Preferably, both the load forecast loss and the electricity price forecast loss are calculated using the mean absolute error function.

[0016] In the second aspect, the power load and electricity price joint forecasting system based on multi-task learning includes: The processor and memory, wherein the memory stores computer program instructions for joint forecasting of power load and electricity price based on multi-task learning, and when the computer program instructions are executed by the processor, the above-described joint forecasting method of power load and electricity price based on multi-task learning is implemented.

[0017] The beneficial effects of this invention are as follows: First, by performing time-series decomposition on historical power load sequences and historical electricity price sequences respectively, this invention constructs a multi-scale joint input feature vector sequence, reducing the difficulty of network learning and improving the model's ability to extract complex features. Second, in the sharing layer, this invention utilizes a learnable channel attention module to generate task-biased shared features, alleviating the task conflict and negative transfer problems existing in traditional hard parameter sharing mechanisms, enabling shared information to serve their respective prediction tasks. Furthermore, this invention utilizes a cross-task interaction gating unit to deeply fuse information between two task-specific networks, uncovering the deep-seated mutual influence relationship between power load and electricity price. Finally, this invention introduces a relationship consistency loss to construct a joint total loss function based on load prediction loss and electricity price prediction loss, ensuring that the prediction results remain consistent with the actual situation in terms of trends and interrelationships. Attached Figure Description

[0018] Figure 1 The flowchart illustrating the steps of the power load and electricity price joint forecasting method based on multi-task learning in this embodiment is shown in the schematic diagram. Figure 2 The schematic diagram illustrates the structural block diagram of the power load and electricity price joint prediction system based on multi-task learning in this embodiment. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] like Figure 1 As shown, the power load and electricity price joint forecasting method based on multi-task learning in this embodiment includes steps S1 to S4: Step S1, obtain historical power load sequence, historical electricity price sequence, meteorological data and calendar data; perform time series decomposition on the historical power load sequence and the historical electricity price sequence respectively to obtain their respective trend components, periodic components and residual components; splice the decomposed components with the meteorological data and calendar data to construct a multi-scale joint input feature vector sequence.

[0021] In this embodiment, the historical power load sequence, the historical electricity price sequence, and the corresponding meteorological data and calendar data of the same period are obtained from the database of the power company or the public data set, with hourly sampling interval. Among them, temperature, humidity and wind speed can be selected as meteorological data; and day of the week and whether it is a holiday can be selected as calendar data.

[0022] In one embodiment, the time series decomposition is performed on the historical power load sequence and the historical electricity price sequence respectively to obtain their respective trend components, periodic components and residual components, including: The decomposition algorithm based on local weighted regression scatter smoothing method is adopted; the historical power load sequence input into the decomposition algorithm is decomposed into load trend component, load periodic component and load residual component; the historical electricity price sequence input into the decomposition algorithm is decomposed into electricity price trend component, electricity price periodic component and electricity price residual component.

[0023] Specifically, the decomposition algorithm based on local weighted regression scatter smoothing method, namely the STL time series decomposition algorithm, is adopted to strip the original time series data into three components that are easy to model and learn through an iterative manner. For example, the historical power load sequence composed of power load data of the past week is decomposed: the power load data of the past week is obtained to constitute a historical power load sequence with a length of The historical power load sequence is input into the STL time series decomposition algorithm, and a smooth long-term trend line is fitted by local regression, the length of the trend line is , which reflects the overall growth or decline of the load in a week; if the trend line is a smooth curve sequence with the value slowly rising from 110 MW to 140 MW, then the sequence is the load trend component. Then, the load trend component is subtracted from the historical power load sequence with a length of to obtain a detrended sequence, and the STL time series decomposition algorithm identifies the repeating patterns in the detrended sequence, such as the daily peak and valley pattern with a unit of twenty-four hours, to form the load periodic component, in which the load is low in the early morning and reaches the peak in the afternoon. The load periodic component is subtracted from the detrended sequence, and after removing the load trend component and the load periodic component, the remaining sequence containing random noise is the load residual component. The historical power load sequence with a length of is decomposed into the load trend component, the load periodic component and the load residual component by the decomposition algorithm, and the lengths of the three components are subsequence.

[0024] Similarly, the same decomposition steps are performed on the historical electricity price series to obtain the electricity price trend component, electricity price cycle component, and electricity price residual component, which have the same length as the historical electricity price series.

[0025] One-hot encoding is performed on the day of the week and whether it is a holiday in the calendar data. At each time step, the load trend component, load cycle component, load residual component, electricity price trend component, electricity price cycle component, electricity price residual component, temperature, humidity, wind speed, and the one-hot encoded day of the week and whether it is a holiday are concatenated along the feature dimension to obtain a multi-scale joint input feature vector, which in turn forms a high-dimensional multi-scale joint input feature vector sequence.

[0026] Step S2: Input the multi-scale joint input feature vector sequence into the shared feature extraction network to obtain the shared feature representation; use the learnable channel attention module to adjust the channel weights of the shared feature representation to generate task-biased shared features.

[0027] In one embodiment, the step of inputting the multi-scale joint input feature vector sequence into a shared feature extraction network to obtain a shared feature representation includes: The shared feature extraction network adopts a bidirectional gated recurrent unit network. The multi-scale joint input feature vector sequence is input into the bidirectional gated recurrent unit network, and temporal features are extracted from both the forward and backward directions through the forward and backward gated recurrent unit layers, respectively. The forward and backward hidden layer states are then concatenated to form a shared feature representation.

[0028] Specifically, the multi-scale joint input feature vector sequence is a time series containing multiple features. The bidirectional gated recurrent unit (BiGRU) network comprises a forward network and a backward network. For an input feature sequence covering the past week (168 hours), i.e., the multi-scale joint input feature vector sequence, the forward network processes data sequentially from the first hour to the 168th hour. At the 168th hour node, it generates a hidden state vector summarizing all past information, resulting in a 128-dimensional vector. Simultaneously, the backward network processes data in reverse chronological order from the 168th hour back to the first hour. At the first hour node, it generates a hidden state vector summarizing all future information, also a 128-dimensional vector, where "future" refers to the future relative to the input multi-scale joint input feature vector sequence. The 128-dimensional vector generated by the forward network and the 128-dimensional vector generated by the backward network are concatenated to form a 256-dimensional shared feature representation. This shared feature representation simultaneously contains a summary of past and future information at a single point in time.

[0029] In an optional embodiment, the shared feature extraction network may also employ a multi-layer long short-term memory network.

[0030] Specifically, the constructed multi-scale joint input feature vector sequence is input into a shared feature extraction network composed of a multi-layer long short-term memory network (LSTM). The LSTM network encodes the temporal features of the input and outputs a shared feature representation containing the common patterns of power load and electricity price.

[0031] In one embodiment, adjusting the channel weights of the shared feature representation using a learnable channel attention module to generate task-biased shared features includes: The shared feature representation is squeezed by a global average pooling layer to obtain channel representations; the channel representations are input into a bottleneck structure consisting of two fully connected layers for activation to learn the weight coefficients of each channel; the learned weight coefficients are multiplied by the shared feature representation by channel to generate task-biased shared features.

[0032] Specifically, the 256-dimensional shared feature representation formed above has 256 channels. This shared feature representation is fed into two independent, learnable channel attention modules. Each channel attention module consists of a global average pooling layer and two fully connected layers, learning channel weights for different tasks. Further, global average pooling is performed on the values ​​of each channel across all time steps, compressing the time-series information of each channel into a single value. For example, when the input shared feature representation has 168 time steps, the 168 values ​​of the first channel are averaged to obtain a single value. The same operation is performed on the second channel until all channels are processed, resulting in a channel representation containing 256 values. This 256-dimensional channel representation is input into the bottleneck structure. A fully connected layer compresses the dimension from 256 to a smaller value, which can be 16. Another fully connected layer restores the dimension to 256, and a sigmoid activation function is used to restrict the output value to between 0 and 1, ultimately generating a 256-dimensional weight coefficient. Each weight coefficient represents the importance of the corresponding channel. The 256-dimensional weight coefficients are multiplied by the original 256-dimensional shared feature representation channel by channel to generate task-biased shared features. High-importance channel features are amplified, while less important channels are suppressed. Furthermore, the channel attention module corresponding to the power load task learns the weight coefficients for each channel specific to the power load task, enhancing beneficial feature channels and suppressing irrelevant or harmful feature channels, thus generating power load task-biased shared features. Similarly, the channel attention module corresponding to the electricity price task learns the weight coefficients for each channel specific to the electricity price task, generating electricity price task-biased shared features.

[0033] Step S3: Input the shared features of the task bias into the dedicated network for power load task and the dedicated network for electricity price task respectively to obtain preliminary power load features and preliminary electricity price features; use the cross-task interaction gating unit to fuse the preliminary power load features and preliminary electricity price features to generate deeply fused load task features and electricity price task features.

[0034] In one embodiment, the step of inputting the shared features of the task bias into the dedicated network for power load tasks and the dedicated network for electricity price tasks, respectively, to obtain preliminary power load features and preliminary electricity price features, includes: The shared features of the task bias are input into the first fully connected layer, which is the dedicated network for power load tasks, and the second fully connected layer, which is the dedicated network for electricity price tasks, respectively, to obtain preliminary power load features and preliminary electricity price features.

[0035] Specifically, the first fully connected layer serves as a dedicated network for power load tasks, and the second fully connected layer serves as a dedicated network for electricity price tasks. The generated shared features biased towards power load tasks are input into the first fully connected layer to extract preliminary power load features. Similarly, the shared features biased towards electricity price tasks are input into the second fully connected layer to extract preliminary electricity price features. Both the preliminary power load features and the preliminary electricity price features can be 128-dimensional.

[0036] In an optional embodiment, both the dedicated network for power load tasks and the dedicated network for electricity price tasks can be composed of a single-layer long short-term memory network (LSTM).

[0037] In one embodiment, the process of fusing preliminary power load characteristics and preliminary electricity price characteristics using a cross-task interaction gating unit to generate deeply fused load task characteristics and electricity price task characteristics includes: generating a load gating signal by passing the preliminary electricity price characteristics through a third fully connected layer and a Sigmoid activation function; performing element-wise multiplication of the load gating signal and the preliminary power load characteristics to obtain load characteristics modulated by electricity price information; generating an electricity price gating signal by passing the preliminary power load characteristics through a fourth fully connected layer and a Sigmoid activation function; performing element-wise multiplication of the electricity price gating signal and the preliminary electricity price characteristics to obtain electricity price characteristics modulated by load information; adding the preliminary power load characteristics and the load characteristics modulated by electricity price information to generate deeply fused load task characteristics; and adding the preliminary electricity price characteristics and the electricity price characteristics modulated by load information to generate deeply fused electricity price task characteristics.

[0038] Specifically, the 128-dimensional preliminary electricity price feature vector is input into a gating unit consisting of a fully connected layer (the third fully connected layer) and a Sigmoid activation function for processing, generating a 128-dimensional load gating signal as an interactive gating weight, with a value between 0 and 1. The load gating signal is then multiplied element-wise with the preliminary power load features to amplify or reduce certain dimensions of the load features, resulting in a load feature modulated by electricity price information. Through residual connections, the load feature modulated by electricity price information is added back to the original preliminary power load features to generate a deeply fused load task feature. This allows the information from the preliminary electricity price features to be integrated into the preliminary power load features, enabling load forecasting to reference electricity price information.

[0039] Similarly, the 128-dimensional preliminary electricity load characteristics are input into a gating unit consisting of a fourth fully connected layer and a Sigmoid activation function for processing, generating a 128-dimensional electricity price gating signal, which also serves as an interactive gating weight; this value is between 0 and 1. The electricity price gating signal is then multiplied element-wise with the preliminary electricity price characteristics to obtain an electricity price characteristic modulated by the load information. This load-modulated electricity price characteristic is then added back to the preliminary electricity price characteristic to generate a deeply fused electricity price task characteristic, which integrates the information from the preliminary electricity load characteristics into the preliminary electricity price characteristic. Through deep fusion of information from two different tasks, each task's feature contains its own original information as well as auxiliary modulation information from the other task.

[0040] Step S4: Based on the deeply fused load task characteristics and electricity price task characteristics, generate the target time's power load forecast value and electricity price forecast value through the prediction head network; calculate the load forecast loss, electricity price forecast loss, and relationship consistency loss, and sum the three by weight to obtain the joint total loss function; update the network parameters based on the joint total loss function.

[0041] In one embodiment, generating the target time's power load forecast and electricity price forecast values ​​through a prediction head network based on the deeply fused load task characteristics and electricity price task characteristics includes: The prediction head network consists of a first multilayer perceptron and a second multilayer perceptron with the same structure. The multilayer perceptron consists of two fully connected layers and a linear output layer. The deeply fused load task features are input into the first multilayer sensor, which outputs the predicted power load value at the target time; the deeply fused electricity price task features are input into the second multilayer sensor, which outputs the predicted electricity price value at the target time.

[0042] Specifically, a first multilayer perceptron, consisting of two fully connected layers and a linear output layer, serves as the power load prediction head network. The deeply fused load task features are input into the first multilayer perceptron to predict the power load. The output dimension is consistent with the prediction step size, thus obtaining the predicted power load value for the future target time. Similarly, a second multilayer perceptron, consisting of two fully connected layers and a linear output layer, serves as the electricity price prediction head network. The deeply fused electricity price task features are input into the second multilayer perceptron to predict the electricity price, generating the predicted electricity price value for the target time. Understandably, taking a 128-dimensional vector as an example of the generated deeply fused load task features, when input into the first multilayer perceptron, the 128-dimensional load task feature vector is transformed into a 64-dimensional intermediate vector through the first fully connected layer, and then transformed into a 32-dimensional vector through the second fully connected layer. During the above transformation process, activation functions are typically used to enhance the model's expressive power. A linear output layer converts the 32-dimensional vector into a single numerical value, which is the predicted power load value for the target time, in MW. Similarly, the generated deep fusion electricity price task feature is also a 128-dimensional vector. When it is input into the second multilayer perceptron, the 128-dimensional electricity price task feature vector is also processed by two fully connected layers for dimensionality reduction. It is then mapped to a single numerical value through a linear output layer. This numerical value is the predicted value of the electricity price at the target time, in yuan per kilowatt-hour.

[0043] Through the above process, it can be ensured that each task can learn the optimal mapping relationship from features to results based on the feature distribution, and map the high-dimensional task features after deep fusion to the prediction results.

[0044] In one embodiment, both the load forecasting loss and the electricity price forecasting loss are calculated using the mean absolute error function.

[0045] The formula for calculating the load forecast loss is as follows: ; In the formula, For load forecasting losses, , The first The actual and predicted power load values ​​for each sample. It is the absolute value symbol. The total number of samples.

[0046] The formula for calculating the electricity price forecast loss is as follows: ; In the formula, Forecasting losses for electricity prices, , The first The actual and predicted electricity prices for each sample. It is the absolute value symbol. The total number of samples.

[0047] Specifically, For the power load, the difference between the predicted and actual power load values ​​is calculated using the mean absolute error function as the load prediction loss. For electricity prices, the difference between the predicted and actual electricity prices is calculated using the mean absolute error function (MAE) as the electricity price prediction loss. For example, the actual electricity load values ​​for three samples are 150MW, 200MW, and 180MW, while the predicted electricity load values ​​at the target time are 155MW, 190MW, and 182MW. The absolute values ​​of their differences are 5, 10, and 2, respectively. The average of these three values ​​is 5.67, and this average value is the load prediction loss for this batch of samples. Similarly, the electricity price prediction loss can be calculated using the exact same method.

[0048] In one embodiment, the method for calculating the relationship consistency loss is as follows: Calculate the Pearson correlation coefficient between the predicted power load series and the predicted electricity price series; calculate the Pearson correlation coefficient between the actual power load series and the actual electricity price series; and take the absolute value of the difference between the two Pearson correlation coefficients as the consistency loss.

[0049] The formula for calculating the relationship consistency loss is as follows: ; In the formula, For relationship consistency loss, For the predicted power load sequence, For the predicted electricity price series, For the predicted power load sequence With the predicted electricity price series The Pearson correlation coefficient between them For the actual power load sequence, For the true electricity price series, For the actual power load sequence With the real electricity price series The Pearson correlation coefficient between them It is the absolute value symbol.

[0050] Specifically, the difference between the correlation between predicted sequences and the correlation between actual sequences is calculated, and this difference is taken as the consistency loss. In essence, the Pearson correlation coefficient between the actual power load sequence and the actual electricity price sequence over a given time period is calculated; the Pearson correlation coefficient between the predicted power load sequence and the predicted electricity price sequence over the same time period is also calculated; the consistency loss is the absolute value of the difference between the two Pearson correlation coefficient values. This consistency loss constrains the correlation between the predicted sequences, ensuring that power load and electricity prices are not only numerically reliable but also consistent with reality in their interaction over time, thus improving the overall accuracy and reasonableness of the joint forecast.

[0051] After calculating the load forecasting loss, electricity price forecasting loss, and relationship consistency loss, a joint total loss function is constructed by weighted summation of the three. During model training, the Adam optimizer is used, and the gradient of the joint total loss function with respect to all learnable parameters in the network is calculated using the backpropagation algorithm. The parameters are then updated in the reverse direction of the gradient until the joint total loss function converges.

[0052] This invention also provides a joint forecasting system for power load and electricity price based on multi-task learning. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions for joint forecasting of power load and electricity price based on multi-task learning. When the computer program instructions are executed by the processor, the joint forecasting method for power load and electricity price based on multi-task learning according to the present invention is implemented.

[0053] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0054] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented by computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0055] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0056] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A power load and price joint prediction method based on multi-task learning, characterized in that, The method comprises the following steps: obtaining historical power load sequences, historical electricity price sequences, meteorological data and calendar data; performing time series decomposition on the historical power load sequences and the historical electricity price sequences respectively to obtain trend components, periodic components and residual components of each sequence; concatenating the decomposed components with the meteorological data and the calendar data to construct a multi-scale joint input feature vector sequence; inputting the multi-scale joint input feature vector sequence into a shared feature extraction network to obtain a shared feature representation; adjusting the channel weights of the shared feature representation by using a learnable channel attention module to generate task-biased shared features; inputting the task-biased shared features into a power load task-specific network and an electricity price task-specific network respectively to obtain preliminary power load features and preliminary electricity price features; and fusing the preliminary power load features and the preliminary electricity price features by using a cross-task interaction gating unit to generate deeply fused load task features and electricity price task features; generating power load prediction values and electricity price prediction values at a target time based on the deeply fused load task features and the electricity price task features by using a prediction head network; calculating load prediction loss, electricity price prediction loss and relationship consistency loss, and then weighting and summing the three losses to obtain a joint total loss function; and updating network parameters based on the joint total loss function. 2.The method of claim 1, wherein, The method further comprises the following steps: adopting a decomposition algorithm based on local weighted regression scatter smoothing; decomposing the historical power load sequences input into the decomposition algorithm into load trend components, load periodic components and load residual components; and decomposing the historical electricity price sequences input into the decomposition algorithm into electricity price trend components, electricity price periodic components and electricity price residual components. 3.The method of claim 1, wherein, The method further comprises the following steps: The shared feature extraction network adopts a bidirectional gated recurrent unit network; the multi-scale joint input feature vector sequence is input into the bidirectional gated recurrent unit network to extract time series features from both positive and negative directions by using the forward and backward gated recurrent unit layers, and the hidden layer states of the forward and backward directions are concatenated to form the shared feature representation. 4.The method of claim 1, wherein, The method further comprises the following steps: The method further comprises the following steps: 5.The method of claim 1, wherein, inputting the channel representation into a bottleneck structure composed of two fully connected layers to perform excitation operation and learn the weight coefficients of each channel; and multiplying the learned weight coefficients with the shared feature representation by channel to generate task-biased shared features. The method further comprises the following steps: The shared features of the task bias are respectively input into a first full connection layer as a power load task exclusive network and a second full connection layer as a price task exclusive network to obtain preliminary power load features and preliminary price features. 6.The method of claim 5, wherein, The preliminary power load features and the preliminary price features are fused by using a cross-task interaction gating unit to generate deeply fused load task features and price task features, including: the preliminary price features are input into a third full connection layer and a Sigmoid activation function to generate a load gating signal, and the load gating signal and the preliminary power load features are subjected to element-level multiplication to obtain load features modulated by price information; the preliminary power load features are input into a fourth full connection layer and a Sigmoid activation function to generate a price gating signal, and the price gating signal and the preliminary price features are subjected to element-level multiplication to obtain price features modulated by load information; the preliminary power load features and the load features modulated by price information are added to generate the deeply fused load task features; and the preliminary price features and the price features modulated by load information are added to generate the deeply fused price task features.

7. The method of claim 6, wherein the method is based on multi-task learning. The power load prediction value and the price prediction value at the target time are generated by a prediction head network based on the deeply fused load task features and the deeply fused price task features, including: The prediction head network is a first multilayer perceptron and a second multilayer perceptron with the same structure, and each multilayer perceptron is composed of two full connection layers and a linear output layer; The deeply fused load task features are input into the first multilayer perceptron to output the power load prediction value at the target time, and the deeply fused price task features are input into the second multilayer perceptron to output the price prediction value at the target time. 8.The method of claim 1, wherein, The calculation method of the relationship consistency loss is: Pearson correlation coefficients between the predicted power load sequence and the predicted price sequence are calculated, and Pearson correlation coefficients between the real power load sequence and the real price sequence are calculated; The absolute value of the difference between the two Pearson correlation coefficients is taken as the relationship consistency loss. 9.The method of claim 1, wherein, Both the load prediction loss and the price prediction loss are calculated by using a mean absolute error function.

10. A power load and price joint forecasting system based on multi-task learning, characterized in that, It includes: A processor and a memory, the memory storing computer program instructions for multi-task learning based power load and price joint prediction, when the computer program instructions are executed by the processor, the multi-task learning based power load and price joint prediction method according to any one of claims 1-9 is realized.

Images