Energy-load-price multi-task joint forecasting method and system for integrated energy systems
Through multi-task learning and feature sharing mechanism, the problem of joint prediction of source and load price in the integrated energy system of traditional prediction methods is solved, and more efficient uncertain prediction is achieved, improving prediction accuracy and model generalization capabilities.
Patent Information
- Application Number
- PCT/CN2024/089699
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-06
- Filing Date
- 2024-04-25
- Publication Date
- 2024-11-14
AI Technical Summary
Traditional comprehensive energy system prediction methods are difficult to achieve accurate joint prediction of source and load price in the context of multiple uncertainties, especially in the case of complex coupling characteristics at the energy-price level, independent prediction efficiency is low and accurate prediction is difficult.
Using multi-task learning method, the training data set is constructed through cross-correlation and autocorrelation analysis, the channel attention and timing attention mechanism are used to extract feature, and the feature sharing is performed through parameter sharing learning mechanism, combining hard sharing and soft sharing mechanisms Make predictions to achieve joint prediction of source and load prices.
It improves the accuracy and efficiency of uncertainty prediction in the comprehensive energy system, enhances the generalization ability of the model, can better extract and classify spatial and temporal coupled features, and improves prediction accuracy.
Smart Images

Figure CN2024089699_14112024_PF_FP_ABST
Abstract
Description
A multi-task joint prediction method and system for source-load price in integrated energy system
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS
[0002] This invention claims priority to Chinese patent application No. 202310519941.5 filed with the State Intellectual Property Office of China on May 6, 2023, entitled “A Multi-task Joint Prediction Method and System for Source-Load Price in an Integrated Energy System”, the entire contents of which are incorporated by reference into this invention and constitute a part of this invention for all purposes. Technical Field
[0003] The present invention belongs to the technical field of integrated energy system planning, and relates to a multi-task joint prediction method and system for source-load price of an integrated energy system. Background Art
[0004] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0005] The penetration of renewable energy sources such as wind power and photovoltaics on the energy supply side of integrated energy systems is gradually increasing. On the energy consumption side, flexible loads such as electric vehicles and energy storage batteries are being connected in large numbers. Furthermore, more private capital is being encouraged to enter the energy market and participate in competitive transactions. This has significantly increased the uncertainty factors in the system, covering multiple links such as energy generation, trading, and consumption. Accurate uncertainty prediction is a prerequisite for ensuring the efficient operation of integrated energy systems.
[0006] Traditional forecasting methods often independently predict three types of uncertainty: renewable energy power on the energy supply side, load on the energy consumption side, and energy prices on the market side. Simple model methods are also relatively mature, using time series analysis or machine learning methods to obtain point forecasts, interval forecasts, and probabilistic forecasts. However, there are complex coupling relationships between multiple random variables in an integrated energy system. In this context, it is crucial to comprehensively consider and analyze the correlations between random variables and their influencing factors, and fully extract their spatial and temporal coupling characteristics to achieve accurate predictions in multivariate situations. Compared to single-variable independent forecasting methods, conducting joint forecasting of multiple random variables can further explore the inherent connections between uncertain factors in an integrated energy system and improve forecasting accuracy and efficiency.
[0007] In short, facing the complex coupling of uncertainties at the energy and price level in integrated energy systems, traditional independent forecasting methods face low efficiency and difficulty in accurate prediction. Further research on the interplay between user behavior, load demand response, and market-level energy prices, and the development of appropriate forecasting models, will be required to achieve accurate joint forecasting of source and load prices in integrated energy systems.
[0008] Summary of the Invention
[0009] In order to solve the above problems, the present invention proposes a multi-task joint prediction method and system for source-load price of an integrated energy system. The present invention can perform correlation analysis on the two uncertainties at the energy level and the price level and carry out joint prediction under the background of multiple uncertainties in the integrated energy system.
[0010] According to some embodiments, the present invention adopts the following technical solutions:
[0011] A multi-task joint prediction method for source-load price of an integrated energy system includes the following steps:
[0012] The spatiotemporal coupling characteristics of the various uncertainty factors affecting the integrated energy system are analyzed from the perspectives of mutual correlation and autocorrelation;
[0013] Construct a training data set and perform feature extraction based on the two correlation analysis results;
[0014] The extracted features are re-extracted using the channel attention and temporal attention mechanisms;
[0015] The parameter sharing learning mechanism is used to share features of the data after secondary extraction. The multi-task learning method takes price prediction as the main task and source and load prediction as the auxiliary tasks. The coupling between the corresponding loads of various tasks adopts a hard sharing mechanism, and the soft sharing mechanism is used between different categories of tasks. The auxiliary tasks share information with the main task to obtain a joint prediction result.
[0016] As an optional implementation, the Pearson product-moment correlation coefficient, the Spearman rank correlation coefficient, and the Kendall rank correlation coefficient are used to perform correlation analysis among multiple loads, renewable energy generation, energy prices, time, and meteorological factors, respectively.
[0017] As an optional implementation method, the autocorrelation coefficient is used to analyze the time series characteristics of each factor to determine the time length of the data used for prediction.
[0018] As an optional implementation, the specific process of extracting features from the two correlation analysis results includes: extracting features from the input data through a convolutional layer, wherein the convolutional layer regularly traverses the input data, performs matrix element multiplication on the input data, and superimposes the deviation amount;
[0019] The feature map output by the convolutional layer is passed to the pooling layer using the activation of the ReLU function. The window of the pooling layer slides over all areas of the input according to the stride size, traverses each position through the window and calculates the output;
[0020] The above process is iterated to obtain the final output features.
[0021] As an optional implementation method, the specific process of secondary extraction of the extracted features using the channel attention and temporal attention mechanisms includes: connecting the channel attention module and the temporal attention module in series, inputting the features through the channel attention module to obtain the channel attention matrix, multiplying the matrix by the original image to obtain the features and serving as the input of the temporal attention module, obtaining the time attention matrix through the temporal attention module, and then multiplying the time attention matrix by the original image to obtain the output features.
[0022] As an optional implementation, the processing process of the channel attention module includes passing the input feature sequence through two parallel MaxPool layers and AvgPool layers, compressing the feature sequence dimension, then expanding the number of channels after compression back to the original number of channels, and obtaining two activated results through the ReLU activation function; adding these two output results element by element, and obtaining the channel attention matrix through the sigmoid activation function.
[0023] The processing process of the temporal attention module includes: splicing the data of different channels according to the time dimension of the feature sequence obtained by the channel attention module to obtain a feature sequence; extracting features through a one-dimensional convolution layer, transforming the dimension of the feature sequence to the original dimension, and then obtaining a temporal attention matrix through the activation function Sigmoid.
[0024] As an optional implementation method, the LSTM network is used for feature sharing. The thermal power load is directly hard-shared through the LSTM network. The wind power, load characteristics and electricity price characteristics are soft-shared between different LSTM network layers through weighted summation. The auxiliary task shares information with the main task during output.
[0025] A multi-task joint prediction system for source-load price of integrated energy system, including:
[0026] A correlation analysis module is configured to analyze the spatiotemporal coupling characteristics of various uncertainty factors affecting the integrated energy system from the perspectives of mutual correlation and autocorrelation;
[0027] A feature extraction module is configured to extract features from the two correlation analysis results;
[0028] The secondary extraction module is configured to perform secondary extraction on the extracted features using channel attention and temporal attention mechanisms;
[0029] The feature sharing module is configured to construct a two-layer feature sharing model using hard sharing and soft sharing mechanisms to perform feature sharing on the data after secondary extraction. The two-layer feature sharing model is divided according to the type of prediction tasks. The coupling between the corresponding loads of each type of task adopts a hard sharing mechanism, and the coupling between different types of tasks adopts a soft sharing mechanism. The auxiliary tasks share information with the main task to obtain a joint prediction result.
[0030] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the steps in the method.
[0031] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the steps in the described method.
[0032] Compared with the prior art, the present invention has the following beneficial effects:
[0033] In the context of multiple uncertainties in integrated energy systems, this paper analyzes the correlations of three major categories of uncertainty—source, load, and price—and conducts joint prediction. The proposed multi-task learning model classifies prediction targets based on correlation and attribute type. In a two-layer feature sharing module, different sharing mechanisms are employed based on the differences in task types, distinguishing between outer and inner layers. This promotes the rational sharing of important features, improves the model's generalization capabilities, and enables efficient and accurate joint prediction of multiple uncertainties. The proposed deep learning network can classify and extract spatial and temporal coupled features in the data, highlighting key features and further improving the model's prediction accuracy.
[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0036] FIG1 is a technical route of a joint prediction method taking into account multiple uncertainties in the present invention;
[0037] FIG2 is a diagram showing example data of the present invention;
[0038] Figure 3 (a)-(c) is a graph showing the results of the cross-correlation analysis in the present invention;
[0039] FIG4 is a graph showing the results of autocorrelation analysis in the present invention;
[0040] FIG5 is a schematic diagram of a feature extraction layer in the present invention;
[0041] FIG6 is a structural diagram of the temporal convolutional attention module in the present invention;
[0042] FIG7 is a schematic diagram of a channel attention module of a temporal convolutional attention module in the present invention;
[0043] FIG8 is a schematic diagram of a temporal attention module of a temporal convolutional attention module in the present invention;
[0044] FIG9 is a schematic diagram of a feature sharing layer in the present invention;
[0045] FIG10 is a comparison diagram of the prediction results of the model in the present invention;
[0046] FIG11 is a comparison diagram of the MAPE probability distribution of the model prediction results in the present invention;
[0047] FIG12 is a comparison diagram of the RMSE probability distribution of the prediction results of the model in the present invention. DETAILED DESCRIPTION
[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0049] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.
[0050] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0051] As shown in Figure 1, addressing multiple prediction objectives and their complex coupling characteristics, this two-layer multi-task learning model vertically categorizes tasks into three main categories: new energy forecasting, multi-load forecasting, and price forecasting, combining commonalities and differences across tasks. The network model also divides the feature sharing model into inner and outer layers based on task type. Appropriate sharing mechanisms are then employed across different layers to achieve feature sharing across tasks under complex coupling conditions, enhancing the model's generalization. Horizontally, the model employs a CNN-SCAM-LSTM network. First, a CNN extracts features from the entire input data. The proposed SCAM is then applied as an attention mechanism to perform feature classification based on task type and extract key features by mining the spatiotemporal characteristics of the data. Finally, an LSTM is used to construct a feature sharing model. By distinguishing between the inner and outer layers, important auxiliary coupling information is further extracted. This allows for hierarchical and targeted shared learning, improving model prediction accuracy.
[0052] It should be noted that the data used in this example comes from the European Electricity System Open Source Dataset, which includes Danish wind power, electro-thermal load, electricity prices, and weather data from 2010 to 2013. To account for the impact of installed wind power capacity on wind power forecasts, the wind power data has been adjusted based on different capacity ratios. The data for wind power, electro-thermal load, and electricity prices from 2010 to 2011 are shown in Figure 2.
[0053] Conduct correlation analysis of the data.
[0054] Cross-correlation analysis
[0055] In integrated energy systems, multiple loads, renewable energy generation, energy prices, time, and meteorological factors interact with each other, and due to different coupling mechanisms and temporal uncertainty, they exhibit complex correlations. To analyze the influencing mechanisms between various factors and the forecast object, fully explore the relevant information in multivariate data, effectively select the input variables of the joint forecast model, and construct the corresponding forecast network structure, quantitative analysis of the dataset is necessary.
[0056] The present invention uses Pearson product-moment correlation coefficient, Spearman rank correlation coefficient and Kendall rank correlation coefficient to explore the correlation between various influencing factors and prediction targets.
[0057] 1. Pearson product-moment correlation coefficient
[0058] The Pearson correlation coefficient measures the linear correlation between two variables, X and Y. It has values between +1 and -1, where 1 is a total positive linear correlation, 0 is a nonlinear correlation, and -1 is a total negative linear correlation. A key mathematical property of the Pearson correlation coefficient is that it is invariant under separate changes in the location and scale of the two variables.
[0059] The calculation formula for the Pearson product-moment correlation coefficient is as follows:
[0060] 2.Spearman rank correlation coefficient
[0061] The Spearman correlation coefficient uses the ranks of two variables to perform linear correlation analysis. It does not require the distribution of the original variables and is a nonparametric statistical method. Therefore, its scope of application is much wider than the Pearson correlation coefficient. The Spearman correlation coefficient can be calculated even if the original data is ranked data. The Spearman correlation coefficient can also be calculated for data that follows the Pearson correlation coefficient, but its statistical power is lower than that of the Pearson correlation coefficient. If there are no duplicate values in the data and the two variables are completely monotonically correlated, the Spearman correlation coefficient will be +1 or -1.
[0062] For sample data with a sample size of n, the calculation formula for the Spearman rank correlation coefficient is as follows:
[0063] Where, d i For data X i and Y i The rank of a number is the position of the number in a column of numbers that it is in, after sorting them from smallest to largest. If some of the values are the same, then the arithmetic average of their positions is taken.
[0064] 3. Kendall rank correlation coefficient
[0065] The Kendall correlation coefficient is also a rank correlation coefficient, which is used to reflect the correlation of categorical variables. It is applicable to the case where both variables are ordered categories, and its value is represented by the Greek letter τ.
[0066] The calculation formula for the Kendall rank correlation coefficient is as follows:
[0067] Where C represents the number of pairs of elements in sample data X and Y that are consistent (two elements form a pair); D represents the number of pairs of elements in sample data X and Y that are inconsistent.
[0068] In general, the Pearson correlation coefficient is a statistic for interval variables, the Spearman correlation coefficient is a statistic for ordinal variables, and the Kendall correlation coefficient is a statistic for categorical variables. The results of the example analysis are shown in Figures 3(a)-3(c). The results of the cross-correlation analysis provide insights into how to select appropriate influencing factors as the input dataset for the prediction network. It can be seen that the correlations between the various prediction targets and influencing factors are consistent under different correlation coefficient metrics. The correlations and degrees of correlation are as follows: electric load has a moderate correlation with thermal load and electricity price, and a weak correlation with temperature and irradiance. Thermal load exhibits a strong negative correlation with temperature and, due to the influence of thermal-electric coupling, a moderate correlation with electric load. Electricity price has a moderate correlation with electric load, consistent with the supply-demand relationship between load and price, and a weak correlation with wind power on the power supply side. Wind power has a strong correlation with wind speed, primarily influenced by weather factors.
[0069] Autocorrelation analysis
[0070] Cross-correlation analysis cannot intuitively reflect the changing patterns of each prediction target at the time level, which makes the determination of the time series length of the input data lack a theoretical basis. This will lead to the following problems. If the time series is too long, it will cause feature redundancy, causing the model to learn many unnecessary parameters. This not only increases the computational load of the model, but also reduces the prediction accuracy of the model on the test set, causing the model to overfit. Too short is not enough to make high-precision predictions for highly nonlinear time series. To this end, the autocorrelation coefficient (ACF) can be used to analyze the time series characteristics of the data. The calculation formula of the autocorrelation coefficient is:
[0071] Where k represents the delay time, represents the mean, and n represents the total time length.
[0072] Figure 4 shows the correlation between electricity load, thermal load, electricity price, and wind power data after a two-week delay. It can be seen that electricity load exhibits both daily and weekly periodicity; thermal load exhibits daily periodicity; electricity load exhibits daily correlation and slightly weaker weekly correlation; and wind power has no significant temporal characteristics. Based on these analysis results, a 24-hour or 24 / 7-hour data length is more appropriate.
[0073] Based on the correlation analysis results, CNN network is used for feature extraction.
[0074] The feature extraction layer structure used in the present invention is shown in Figure 5. First, the input data is subjected to feature extraction through the convolution layer, and its core operation is the convolution kernel. The convolution kernel will regularly traverse the input data, perform matrix element multiplication and summation on the input data within its receptive field, and superimpose the deviation. Then, the feature map output by the convolution layer is passed to the pooling layer after activation of the ReLU function. The pooling layer belongs to the downsampling process and is mainly used for data dimensionality reduction to avoid overfitting. Similar to the convolution layer, the pooling layer operator consists of a fixed-size window, which also slides across all areas of the input according to the stride size, traverses each position through the window and calculates the output. Finally, iterate according to the aforementioned structure to obtain the final output feature.
[0075] Important features are further extracted based on the attention module.
[0076] The attention mechanism model used in this paper is mainly improved with reference to the Convolutional Block Attention Module (CBAM), and a Sequential Convolution Attention Module (SCAM) is studied. The SCAM model studied in this paper is shown in Figure 6. The overall process structure is similar to that of CBAM, consisting of a Channel Attention Module (CAM) and a Sequential Attention Module (SAM) connected in series.
[0077] The model formula is expressed as:
[0078] In the formula, the input feature F∈R C*T , and then the channel attention matrix M is obtained through the channel attention module c ∈R C*1 , the matrix M c Multiply the original image to obtain the feature F′∈R C*T And as the input of the temporal attention module, the temporal attention matrix M is obtained through the temporal attention module t ∈R C*T , and then the matrix M t Multiply with the original image to obtain the output feature F″∈R C*T .
[0079] The channel attention module in the proposed SCAM model maintains the same overall structure as that in CBAM, primarily adjusting the data and layer dimensions. The model structure is shown in Figure 7. First, the input feature sequence passes through two parallel MaxPool and AvgPool layers, compressing the feature sequence dimension from C*T to C*1. Next, it passes through the Share MLP module, where the number of channels is first compressed to 1 / r times the original number (reduction rate), then expanded back to the original number of channels. Two activated results are obtained through the ReLU activation function. Finally, these two output results are element-wise added and the channel attention matrix is obtained through the sigmoid activation function.
[0080] The expression of the above model is:
[0081] Figure 8 shows the architecture of the temporal attention module in the proposed SCAM model. First, the feature sequence generated by the channel attention module is concatenated along the time dimension using the Concat operation to concatenate data from different channels, resulting in a feature sequence of dimension 1*(C*T). Feature extraction is then performed using a one-dimensional convolutional layer with a kernel size of 3, and padding is set to 3 to maintain the dimension of the feature sequence. Finally, a view operation is used to transform the feature sequence to C*T, and then a Sigmoid activation function is applied to obtain the temporal attention matrix.
[0082] The expression of the above model is:
[0083] Feature sharing based on LSTM network.
[0084] After important features are extracted by SCAM, the data is passed to the feature sharing layer. This part mainly considers the correlation between prediction targets and shares information with the features extracted by the previous network. The focus is on selecting a suitable network structure and feature sharing mechanism.
[0085] The source-load-price joint prediction model constructed in this invention features numerous parameters and a complex structure, making it less susceptible to overfitting and exhibiting enhanced model generalization. In the feature sharing component of multi-task learning, the feature sharing model is divided into inner and outer layers according to the task type, employing different sharing mechanisms. Feature sharing between the three major prediction tasks of source, load, and price constitutes the outer layer. Given the complex correlations and different influencing mechanisms, a soft sharing mechanism is employed. Feature sharing between subtasks within each task constitutes the inner layer. For example, in load forecasting, where multiple loads are highly coupled, feature sharing between different loads constitutes the inner layer, employing a hard sharing mechanism. Furthermore, considering the impact of source and load on price, a two-layer feature sharing model is constructed, with price prediction as the primary task and source and load prediction as auxiliary tasks. The specific model for this example is shown in Figure 9. Thermal power loads are directly hard-shared via an LSTM network. Wind power, load characteristics, and electricity price characteristics are soft-shared across different LSTM network layers through weighted summation. The auxiliary tasks share information with the primary task at the output of the fully connected layer.
[0086] Example results analysis
[0087] The construction and training of the source-load-price dual-layer joint prediction model in this paper were conducted within the PyTorch deep learning framework. The hardware platform used an Intel Core i7 CPU, and the computational data was derived from Danish data from 2010 to 2012. The training and validation sets used data from 2010 to 2011, while the test set used data from 2012. The model predicted electricity and heat load, wind power, and electricity prices with a 24-hour step size.
[0088] This example compares the prediction results of the proposed CNN-SCAM-LSTM-MTL model with those of the single-task CNN-LSTM model, the multi-task CNN-LSTM-MTL model, and the CNN-CBAM-LSTM-MTL model to verify the effectiveness of the designed two-layer multi-task joint prediction model.
[0089] As shown in Figure 10, prediction results for the week from January 2nd to 8th in the test set were compared. It can be seen that in the comparison between single-task learning and multi-task learning, the prediction performance of the CNN-LSTM-MTL model alone is inferior to that of single-task learning. However, the prediction results of the CNN-CBAM-LSTM-MTL and CNN-SCAM-LSTM-MTL models using the attention mechanism are comparable to those of the single-task learning CNN-LSTM model and are closer to the actual values. This is because the correlation between the prediction targets in the dataset used in this example is moderate. Using only the input features in a multi-task learning framework, which influences each other, places higher demands on the extraction of important features. This result confirms the necessity of using the attention mechanism.
[0090] This example predicts the entire year's data from the test set and uses Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) as model evaluation metrics to calculate the day-ahead forecast accuracy. The results are compared in Tables 1 and 2. It can be seen that the CNN-LSTM-MTL network, using only a multi-task learning framework for joint forecasting, has difficulty extracting important features tailored to the forecast target, resulting in inferior forecasting performance compared to the CNN-LSTM model. However, the addition of the attention mechanism significantly improves the forecasting performance of the CNN-CBAM-LSTM-MTL and CNN-SCAM-LSTM-MTL models. This demonstrates that the attention mechanism is capable of extracting and classifying important features.
[0091] Comparing the CNN-LSTM model, it can be seen that the CNN-CBAM-LSTM-MTL model's prediction accuracy for electric and thermal loads is still lower than that of single-task learning, but the prediction effects for wind power and electricity prices are better. This is because electric and thermal loads have significant periodicity, while wind power and electricity prices are more random and volatile, and the temporal characteristics of the data are different from the time dimension. Although CBAM can extract important features from the spatial dimension (i.e., the channels in the feature data), it superimposes the feature data of different channels from the time dimension to obtain a common attention weight matrix. This makes CBAM unable to extract the differences in temporal features in the instance data. For this reason, the present invention studies SCAM.
[0092] The CNN-SCAM-LSTM-MTL model studied in this paper achieves better overall prediction accuracy than both the CNN-LSTM model and the CNN-CBAM-LSTM-MTL model. This is because the SCAM model studied in this paper addresses the shortcomings of CBAM in extracting time series features, better balances the spatial and temporal characteristics of feature information, and decouples input features from both spatial and temporal dimensions according to the prediction target, thereby improving prediction accuracy. This validates the effectiveness of the multi-task learning-based source-charge-price dual-layer joint prediction model and SCAM studied in this paper.
[0093] Table 1 Comparison of MAPE indicators
[0094] Table 2 Comparison of RMSE indicators
[0095] This example analyzes the probability distribution of prediction accuracy using the MAPE and RMSE metrics for 2012 on a daily basis. The probability density functions are shown in Figures 11 and 12, and the confidence intervals at a 90% confidence level are shown in Tables 3 and 4. Overall, the CNN-SCAM-LSTM-MTL network achieves superior prediction accuracy peaks and confidence intervals, demonstrating the stability of the prediction results. This further demonstrates that the multi-task learning model studied in this paper has improved prediction accuracy and can improve model generalization by leveraging training information across multiple tasks.
[0096] Table 3 Comparison of confidence intervals of MAPE indicators
[0097] Table 4 Comparison of RMSE confidence intervals
[0098] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0099] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.
[0100] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.
[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.
[0102] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.
[0103] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.
Claims
1. A multi-task joint prediction method for source-load price of integrated energy system, characterized by: The following steps are involved: The spatiotemporal coupling characteristics of the various uncertainties affecting the integrated energy system are analyzed from the perspectives of mutual correlation and autocorrelation; Feature extraction is performed on the two correlation analysis results; The extracted features are re-extracted using the channel attention and temporal attention mechanisms; The parameter sharing learning mechanism is used to share the features of the data after secondary extraction. The multi-task learning method takes price prediction as the main task and source and load prediction as the auxiliary task. The coupling between the corresponding loads of various tasks adopts a hard sharing mechanism, and the soft sharing mechanism is used between tasks of different categories. The auxiliary tasks share information with the main task to obtain the joint prediction results.
2. A multi-task joint prediction method for source-load price of integrated energy system as claimed in claim 1, characterized in that: The Pearson product-moment correlation coefficient, Spearman rank correlation coefficient and Kendall rank correlation coefficient were used to analyze the correlations among multiple loads, renewable energy generation, energy prices, time and meteorological factors.
3. A multi-task joint prediction method for source-load price of integrated energy system as claimed in claim 1, characterized in that: The autocorrelation coefficient is used to analyze the time series characteristics of each factor and determine the time length of the data used for prediction.
4. A multi-task joint prediction method for source-load price of integrated energy system as claimed in claim 1, characterized in that: The specific process of extracting features from the two correlation analysis results includes: extracting features from the input data through a convolution layer, the convolution layer regularly traverses the input data, performs matrix element multiplication and summation on the input data, and superimposes the deviation amount; The feature map output by the convolutional layer is passed to the pooling layer using the activation of the ReLU function. The window of the pooling layer slides over all regions of the input according to the stride size, traverses each position through the window and calculates the output. The above process is iterated to obtain the final output features.
5. A multi-task joint prediction method for source-load price of integrated energy system as claimed in claim 1, characterized in that: The specific process of secondary extraction of the extracted features using the channel attention and temporal attention mechanisms includes: connecting the channel attention module and the temporal attention module in series, inputting the features through the channel attention module to obtain the channel attention matrix, multiplying the matrix by the original image to obtain the features and using it as the input of the temporal attention module, obtaining the time attention matrix through the temporal attention module, and then multiplying the time attention matrix by the original image to obtain the output features.
6. A multi-task joint prediction method for source-load price of an integrated energy system as claimed in claim 1 or 5, characterized in that: The processing process of the channel attention module includes passing the input feature sequence through two parallel MaxPool layers and AvgPool layers, compressing the feature sequence dimension, and then expanding the number of channels back to the original number of channels after compression, and obtaining two activated results through the ReLU activation function; Add these two output results element by element and use the sigmoid activation function to get the channel attention matrix; or, The processing process of the temporal attention module includes: splicing the data of different channels according to the time dimension of the feature sequence processed by the channel attention module to obtain a feature sequence; The features are extracted through a one-dimensional convolutional layer, and then the dimension of the feature sequence is transformed into the original dimension, and then the temporal attention matrix is obtained through the activation function Sigmoid.
7. A multi-task joint prediction method for source-load price of integrated energy system as claimed in claim 1, characterized in that: The LSTM network is used for feature sharing. The thermal power load is directly hard-shared through the LSTM network. The wind power, load characteristics and electricity price characteristics are soft-shared through weighted sums between different LSTM network layers, and the auxiliary task shares information with the main task when outputting.
8. A multi-task joint prediction system for source-load price of integrated energy system, characterized by: include: A correlation analysis module is configured to analyze the spatiotemporal coupling characteristics of various uncertainty factors affecting the integrated energy system from the perspectives of mutual correlation and autocorrelation; A feature extraction module is configured to extract features from two correlation analysis results; The secondary extraction module is configured to perform secondary extraction on the extracted features using channel attention and temporal attention mechanisms; The feature sharing module is configured to build a two-layer feature sharing model using hard sharing and soft sharing mechanisms to share features on the data after secondary extraction. The two-layer feature sharing model is divided according to the prediction task type. The coupling between the corresponding loads of each type of task adopts a hard sharing mechanism, and the soft sharing mechanism is adopted between tasks of different categories. The auxiliary task shares information with the main task to obtain a joint prediction result.
9. A computer-readable storage medium, characterized in that: A plurality of instructions are stored therein, and the instructions are suitable for being loaded by a processor of a terminal device and executing the steps of the method described in any one of claims 1 to 7.
10. A terminal device, characterized in that: The method comprises a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store a plurality of instructions, wherein the instructions are suitable for being loaded by the processor and executing the steps in the method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Multi-task mass user load prediction method based on multi-channel convolutional neural network
CN113988373A
Comprehensive energy system source load price multi-task joint prediction method and system
CN116432861A
System and method for machine learning architecture for multi-task learning with dynamic neural networks
US20230115113A1
Cited By
Power grid dispatching demand response system and method
CN119518991A
Wind speed prediction super-capacity energy storage grid connection method and system based on input quantity position analysis
CN119628034A
Load data management and control method and system for enterprise platform
CN119740793A
Diet health index prediction method fusing M-One coding and deep learning
CN119763832A
Hierarchical double-model fused wind power prediction data model
CN119830958A
