A method for predicting a thermal load curve by classifying a multiple-width learning attention mechanism
By employing a classification-based multi-width learning attention mechanism, the heat load curves are divided into two categories, and historical data is used to optimize the model. This solves the problems of inaccurate prediction and computational complexity in existing technologies, and achieves more efficient heat load prediction and new energy utilization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGXI UNIV
- Filing Date
- 2022-02-28
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies fail to effectively consider prior knowledge of user classification and historical data in predicting the heat load curves of distributed ground source heat pump storage users, resulting in inaccurate predictions, complex calculations, and long neural network model building times.
A classification multi-width learning attention mechanism is adopted. By classifying the heat load curve into two categories, a corresponding prediction model is established. Prior knowledge is formed using historical data, and the model is optimized by performing two modeling operations. The attention mechanism is then combined to improve the prediction accuracy.
It has enabled more accurate prediction of heat load curves, improved the utilization rate of wind and solar energy, and enhanced the power system's ability to absorb new energy sources.
Smart Images
Figure CN114580277B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-energy complementarity in power systems and relates to an artificial intelligence-based prediction method. It is applicable to the prediction of heat load curves for distributed ground source heat pump storage users in power systems and is also applicable to the consumption of new energy sources in power grids. Background Technology
[0002] During seasons when heat energy is needed, people desire an ample supply of heat. Distributed ground source heat pump storage systems can utilize excess wind and solar energy, storing heat energy as electrical load when wind or solar energy is surplus. Predicting the heat load curves of distributed ground source heat pump storage system users is crucial for heat and electricity dispatch. For households with two or more residents who work outside during the day, peak heat load curves typically occur between 7:00 AM and 9:00 AM and between 7:00 PM and 11:00 PM, with significant curve fluctuations. However, for households with home-based workers or elderly people and children, heating is required throughout the day, but the curve is more stable. Grouping similar curves together is more beneficial for establishing subsequent prediction models.
[0003] The existing methods: (1) cannot consider classification, and the prediction of the heat load curve of distributed ground source heat pump heat storage users is inaccurate; (2) single modeling, and the accuracy is not high; (3) the establishment of prediction models of neural networks (such as deep neural networks and BP neural networks) requires complicated calculations and takes a long time; (4) the historical data of users are not considered as prior knowledge.
[0004] The proposed method's classification multi-width learning attention mechanism inherits the advantages of simple structure and fast computation speed of width learning itself. It also separates the heat load, considers prior knowledge, and improves the prediction accuracy of the proposed method by adding an optimization coefficient matrix in the secondary modeling. It is more suitable for real-time load prediction than deep neural networks and BP neural networks. Summary of the Invention
[0005] This invention proposes a heat load curve prediction method based on a classification multi-width learning attention mechanism. The method improves the utilization rate of wind and solar energy through the following steps:
[0006] Step (1) Give the heat load curves of the distributed ground source heat pump heat storage users for four consecutive days, and divide the heat load curves into two categories: the first category is relatively stable curves and the second category is large fluctuation curves. Establish corresponding heat load prediction models for the two categories of heat load curves respectively.
[0007] Given all heat load curves for any four consecutive days in winter, and categorize all heat load curves for each of these four days into two types: the first type of curves is relatively stable, while the second type of curves fluctuates significantly; in the case of... For each heat load, each heat load curve has 96 time points, each time point lasting 5 minutes. If the following conditions are met:
[0008] (1)
[0009] If so, it belongs to the first category; otherwise, it belongs to the second category. This indicates the absolute value; For the first The heat load in the first Heat load values at each time point This indicates the absolute value; For the first The heat load in the first Heat load values at each time point;
[0010] Step (2) uses historical data from the previous week to form attention mechanism values with prior knowledge, highlighting important elements, and forming the first... The heat load in the first Attention mechanism value at each time point for:
[0011] (2)
[0012] For the first time last week The heat load in the first The heat load values at each time point; softmax() is the softmax activation function;
[0013] Step (3): Use the heat load curve data of the first and second days as the initial training set, the heat load curve data of the second and third days as the enhanced training set, and the heat load curve data of the third and fourth days as the test set.
[0014] The initial training set is used for multi-width learning to build the first model, which is called the first modeling, and the weights obtained in the first modeling process are stored.
[0015] The enhanced training set is then input into the model established in the first step to optimize the model. This is called secondary modeling. In the second modeling, the increased optimization coefficient matrix is stored.
[0016] Then, multiply it by the attention mechanism weights calculated in step (2);
[0017] Secondly, the input data from the test set is input into the established multi-width learning prediction model and the predicted value is obtained. The root mean square error between the predicted value and the true output value in the test set is calculated, and the error of the established model is obtained.
[0018] (3.1) Use the heat load curve data of the first and second days as the initial training set, the heat load curve data of the second and third days as the enhanced training set, and the heat load curve data of the third and fourth days as the test set.
[0019] The first type of heat load is One, the second type of load is indivual, ;
[0020] The dataset representations for the initial training set, enhanced training set, and test set for the first and second types of heat loads are given as follows:
[0021] Category 1:
[0022] Category Two:
[0023] in, and This is the initial training set for the first type of heat load. and For the enhanced training set of the first type of heat load, and This is the test set for the first type of heat load; and This is the initial training set for the second type of heat load. and For the enhanced training set of the second type of heat load, and This is the test set for the second type of heat load;
[0024] (3.2) Input the initial training set Generate the first Group mapping feature node matrix:
[0025] (3)
[0026] in, Generate and store the random weight matrix. Generate and store the random deviation matrix. This represents the number of feature node matrices mapped; sigmoid() is the sigmoid activation function;
[0027] (3.3) The first feature node is generated from the mapping feature node matrix. Group-enhanced node matrix:
[0028] (4)
[0029] in, Generate and store the random weight matrix. Generate and store the random deviation matrix. , , This represents the number of enhanced feature node matrices; tanh() is the tanh activation function;
[0030] (3.4) Find the random weight matrix:
[0031] (5)
[0032] in for and Combination matrix;
[0033] (3.5) Perform secondary modeling, inputting the data from the augmented training set. The input is fed into the trained model above, and the predicted value is calculated based on the stored and calculated weights. The solution process is as follows:
[0034] Update # Group mapping feature node matrix:
[0035] (6)
[0036] Update the first mapping feature node matrix using the updated matrix. Group-enhanced node matrix:
[0037] (7)
[0038] Considering the secondary output prediction value of the attention mechanism:
[0039] (8)
[0040] in Attention mechanism value The row vector formed;
[0041] (3.6) Calculate the optimization coefficient matrix added in the quadratic modeling:
[0042] (9)
[0043] (3.7) Input the test set The final predicted value is obtained by calculating the weights in steps (3.2) to (3.6). This predicted value is then compared with the true value to calculate the root mean square error between the predicted value and the true output value in the test set. This error represents the prediction model error of the multi-width learning attention mechanism.
[0044] (10)
[0045] (3.8) Model the second type of heat load according to steps (3.2) to (3.7) to obtain the prediction model of the heat load curves of the two types of heat loads;
[0046] Step (3) The established multi-width learning attention mechanism model is used to predict the heat load curve of distributed ground source heat pump storage users;
[0047] Step (4): When there is a surplus of wind and solar energy, the power system dispatch center, based on the predicted heat load curve of the distributed ground source heat pump storage users, will use the surplus electrical energy of wind and solar energy as electrical load to store heat energy and supply the heat energy to the distributed ground source heat pump storage users, thereby improving the utilization rate of wind and solar energy and enhancing the power system's ability to absorb wind and solar energy.
[0048] The present invention has the following advantages and effects compared with the prior art:
[0049] (1) The present invention adopts a query-based attention mechanism to form prior knowledge of the time periods of different types of distributed ground source heat pump heat storage users, forming prior knowledge with "memory" that conforms to human activities, and can obtain more accurate heat load curves.
[0050] (2) This invention is a classification prediction, namely, multi-width learning of "classification"; the heat load curve of each day is divided into two categories. The first category curve is relatively stable, while the second category curve fluctuates more. Classifying similar curves into one category is more conducive to the establishment of subsequent prediction models and to improving the accuracy of prediction models.
[0051] (3) This invention involves two modeling processes, namely "multiple" width learning. The initial training set is used for the first model building through multiple width learning, which is called primary modeling. Then, the enhanced training set is input into the model built in the first model to optimize the model, which is called secondary modeling. The optimization of the model through secondary modeling can improve the accuracy of load forecasting.
[0052] (4) The method of the present invention can provide a scheduling basis for the power system dispatch center, realize multi-energy complementarity, thereby improving the utilization rate of wind and solar energy and enhancing the power system's ability to absorb wind and solar energy. Attached Figure Description
[0053] Figure 1 This is a flowchart of the heat load classification method of the present invention.
[0054] Figure 2 This is the training process of the multi-width learning attention mechanism prediction model established by the method of this invention.
[0055] Figure 3 The diagram shows the working principle of the power grid in which the heat load curve of the distributed ground source heat pump heat storage device mentioned in the method of this invention is displayed. Detailed Implementation
[0056] This invention proposes a method for predicting the heat load curve of a classification multi-width learning attention mechanism, which is described in detail below with reference to the accompanying drawings:
[0057] Figure 1 This is a flowchart of the heat load classification method of the present invention. First, the heat load curves of all distributed ground source heat pump storage users are given for a day. Then, it is determined whether the difference between any two adjacent time points of each heat load curve is less than 1 kW. If so, it indicates that the curve is relatively stable and is a heat load curve of the first type of heat load; if not, it indicates that the curve fluctuates greatly and is a heat load curve of the second type of heat load.
[0058] Figure 2 This invention describes the training process of the multi-width learning attention mechanism prediction model established by the method of this invention. First, a model is built using an initial training set to obtain the corresponding weight matrix. Second, the output prediction data is calculated using the weight matrix obtained in the first model and the input data in the augmented training set. Then, the optimized coefficient matrix is calculated using the calculated output prediction data and the actual input data in the augmented training set. The output of the second model is then multiplied by the attention mechanism weights. Finally, the accuracy of the established prediction model is verified using a test set.
[0059] Figure 3 The diagram shows the working principle of the distributed ground source heat pump storage device user's heat load curve in the power grid. When wind and solar energy are in surplus, the distributed ground source heat pump storage device operates as electrical load to store heat energy, which is then supplied to the distributed ground source heat pump storage device user, thus improving the utilization rate of wind and solar energy.
[0060] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for predicting the heat load curve of a classification multi-width learning attention mechanism, characterized in that, The steps of this method to improve the utilization rate of wind and solar energy are as follows: Step (1) Give the heat load curves of the distributed ground source heat pump heat storage users for four consecutive days, and divide the heat load curves into two categories: the first category is relatively stable curves and the second category is large fluctuation curves. Establish corresponding heat load prediction models for the two categories of heat load curves respectively. Given all heat load curves for any four consecutive days in winter, and categorize all heat load curves for each of these four days into two types: the first type of curves is relatively stable, while the second type of curves fluctuates significantly; in the case of... For each heat load, each heat load curve has 96 time points, each time point lasting 5 minutes. If the following conditions are met: (1) If so, it belongs to the first category; otherwise, it belongs to the second category. This indicates the absolute value; For the first The heat load in the first Heat load values at each time point This indicates the absolute value; For the first The heat load in the first Heat load values at each time point; Step (2) uses historical data from the previous week to form attention mechanism values with prior knowledge, highlighting important elements, and forming the first... The heat load in the first Attention mechanism value at each time point for: (2) For the first time last week The heat load in the first The heat load values at each time point; softmax() is the softmax activation function; Step (3): Use the heat load curve data of the first and second days as the initial training set, the heat load curve data of the second and third days as the enhanced training set, and the heat load curve data of the third and fourth days as the test set. The initial training set is used for multi-width learning to build the first model, which is called the first modeling, and the weights obtained in the first modeling process are stored. The enhanced training set is then input into the model established in the first step to optimize the model. This is called secondary modeling. In the second modeling, the increased optimization coefficient matrix is stored. Then, multiply it by the attention mechanism weights calculated in step (2); Secondly, the input data from the test set is input into the established multi-width learning prediction model and the predicted value is obtained. The root mean square error between the predicted value and the true output value in the test set is calculated, and the error of the established model is obtained. (3.1) Use the heat load curve data of the first and second days as the initial training set, the heat load curve data of the second and third days as the enhanced training set, and the heat load curve data of the third and fourth days as the test set. The first type of heat load is One, the second type of load is indivual, ; The dataset representations for the initial training set, enhanced training set, and test set for the first and second types of heat loads are given as follows: Category 1: Category Two: in, and This is the initial training set for the first type of heat load. and For the enhanced training set of the first type of heat load, and This is the test set for the first type of heat load; and This is the initial training set for the second type of heat load. and For the enhanced training set of the second type of heat load, and This is the test set for the second type of heat load; (3.2) Input the initial training set Generate the first Group mapping feature node matrix: (3) in, Generate and store the random weight matrix. Generate and store the random deviation matrix. This represents the number of feature node matrices mapped; sigmoid() is the sigmoid activation function; (3.3) The first feature node is generated from the mapping feature node matrix. Group-enhanced node matrix: (4) in, Generate and store the random weight matrix. Generate and store the random deviation matrix. , , This represents the number of enhanced feature node matrices; tanh() is the tanh activation function; (3.4) Find the random weight matrix: (5) in for and Combination matrix; (3.5) Perform secondary modeling, inputting the data from the augmented training set. The input is fed into the trained model above, and the predicted value is calculated based on the stored and calculated weights. The solution process is as follows: Update # Group mapping feature node matrix: (6) Update the first mapping feature node matrix using the updated matrix. Group-enhanced node matrix: (7) Considering the secondary output prediction value of the attention mechanism: (8) in Attention mechanism value The row vector formed; (3.6) Calculate the optimization coefficient matrix added in the quadratic modeling: (9) (3.7) Input the test set The final predicted value is obtained by calculating the weights in steps (3.2) to (3.6). This predicted value is then compared with the true value to calculate the root mean square error between the predicted value and the true output value in the test set. This error represents the prediction model error of the multi-width learning attention mechanism. (10) (3.8) Model the second type of heat load according to steps (3.2) to (3.7) to obtain the prediction model of the heat load curves of the two types of heat loads; Step (3) The established multi-width learning attention mechanism model is used to predict the heat load curve of distributed ground source heat pump storage users; Step (4): When there is a surplus of wind and solar energy, the power system dispatch center, based on the predicted heat load curve of the distributed ground source heat pump storage users, will use the surplus electrical energy of wind and solar energy as electrical load to store heat energy and supply the heat energy to the distributed ground source heat pump storage users, thereby improving the utilization rate of wind and solar energy and enhancing the power system's ability to absorb wind and solar energy.