Regional new energy output and load prediction method based on source-load common mode

Abstract

The invention discloses a regional new energy output and load prediction method based on a source-load common mode, and the method comprises the following steps: 1), collecting a source-load end historical sequence and same-time-period meteorological data, 2), obtaining a multi-mode feature sequence of an electrical load and a new energy output sequence at different frequencies through an original sequence of variational mode decomposition; 3) fusing the multi-modal feature sequence, the meteorological feature sequence and the time coding feature sequence to obtain a multi-dimensional feature fusion sequence group; 4) constructing a model containing a pyramid attention module based on a multi-way tree structure; and 5) collecting features given by the last node of all scales in the pyramid attention module, mapping the features to all future time steps in the batch, and obtaining final prediction through a full connection layer of a cross-channel dimension. The method has the beneficial effect that the common mode effect of a complex coupling relation under the influence of multiple factors such as a supply side and a demand side is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy output and load forecasting, and in particular to a method for regional new energy output and load forecasting based on a source-load common mode. Background Art

[0002] In the existing source-load forecasting research, the new energy power generation side and the load demand side are mainly modeled separately. In terms of forecast type, the source side mainly forecasts wind turbine output and photovoltaic output; in terms of time scale, it can be divided into ultra-short-term, short-term, medium- and long-term forecasts; in terms of spatial scale, it can be divided into regional-level forecasts and station-level forecasts; the new energy output forecasts at different time and space scales differ in research methods and applications according to the accuracy requirements; the load side mainly forecasts electricity load demand, which can be further divided into regional-level total social electricity consumption forecasts, regional-level annual load peak forecasts, regional-level long-period time series load curve forecasts, and industry electricity load forecasts. In traditional practices, one is to obtain meteorological data based on the numerical weather forecast (NWP) system, use the Pearson correlation coefficient method or other correlation calculation methods to obtain meteorological characteristics that are strongly correlated with renewable energy output and regional load, and combine them with the historical characteristics of renewable energy and historical load characteristics to independently establish a model to predict renewable energy output and regional load; the other is to use the idea of ​​time series analysis, using empirical mode decomposition (EMD), singular spectrum analysis (SSA), STL decomposition, factor decomposition method, SEATS decomposition method, etc., to decompose the original time series into long-term trends, seasonal trends, periodic fluctuations, random noise and other parts, analyze the characteristics of renewable energy output and load, and then use probability statistics methods, machine learning methods, deep reinforcement learning methods, etc. to establish a prediction model. The above analysis methods and ideas still have the following needs and challenges in solving accurate regional power load forecasts and renewable energy output forecasts:

[0003] (1) There are many influencing factors. The source load data itself is highly volatile and random, and is affected by many factors such as geographical location, seasonal changes, weather, social development, policy planning, etc., making it difficult to consider comprehensively;

[0004] (2) The independent models are disconnected in time and space. The independent wind power output forecast model, photovoltaic output forecast model, and load forecast model fail to fully utilize the temporal and spatial correlation information, and fail to consider the combined effects of meteorological changes and different modes on both the supply and demand sides;

[0005] (3) The model has weak long-term learning ability. There is a lack of adaptive feature learning methods. By manually combining features, due to the differences between modalities, it is not possible to better refer to different dimensions and more hidden information during long-term prediction, which affects the model's prediction ability. Summary of the invention

[0006] The present invention aims to overcome the deficiencies of the above-mentioned prior art and provides a method for forecasting regional renewable energy output and load based on source-load common mode, which is specifically implemented by the following technical solutions:

[0007] The method for predicting regional new energy output and load includes the following steps:

[0008] Step 1) Collect the historical sequence of the source and load end and the meteorological data of the same period,

[0009] Step 2) obtaining the multi-modal characteristic sequences of the power load and the new energy output sequence at different frequencies through the original sequence of variational mode decomposition;

[0010] Step 3) Decomposing the power load and new energy output sequences by using different time scale features to obtain feature sequences of different period lengths in the time dimension; calculating the meteorological data of the corresponding area and the corresponding load sequence and the new energy output sequence by using the Pearson correlation coefficient method, selecting meteorological features with larger correlation coefficients, and combining them into meteorological feature sequences; performing unique hot encoding on the time labels of the time dimension to obtain a time coding feature sequence; fusing the multimodal feature sequence, the meteorological feature sequence, and the time coding feature sequence to obtain a multidimensional feature fusion sequence group;

[0011] Step 4) constructing a model containing a pyramid attention module based on a multi-tree structure, and optimizing the model based on set rules; the pyramid attention module constructs a multi-resolution representation of the original sequence at different time scales, and the C-tree of the pyramid structure is composed of nodes of different time scales;

[0012] Step 5) Collect the features given by the last node of all scales in the pyramid attention module, map them to all future time steps in the batch, obtain the final prediction through a fully connected layer across the channel dimension, and output all future prediction results linearly together.

[0013] The method for forecasting regional renewable energy output and load is further designed in that the model further includes: a data embedding module, which embeds the acquired information including observation data, covariate data and location respectively, and then adds various parameters in the information together to form a new feature representation;

[0014] The multi-time resolution C-tree module uses a coarse scale to construct a multi-resolution C-tree and completes the information transfer between different scales through nodes.

[0015] The further design of the regional new energy output and load forecasting method is that the variational mode decomposition in step 2) specifically includes the following steps:

[0016] Step 2-1) Initialize the modal function set Center frequency set Lagrangian operator λ 1 , initial iteration number b = 0;

[0017] Step 2-2) Set the number of iterations b = b + 1;

[0018] Step 2-3) When w ≥ 0 is satisfied, update the spectrum of each mode

[0019]

[0020] Where, represents the penalty coefficient, which determines the bandwidth of the component mode function; w represents the center frequency, w k represents the center frequency of the kth mode; The Lagrangian operator representing the signal to be decomposed; The frequency domain representation of the signal to be decomposed updates the center frequency w k :

[0021]

[0022] Step 2-4) Update the Lagrange multiplier λ:

[0023]

[0024] Among them, ε is the noise tolerance parameter;

[0025] Step 2-5) Repeat the above steps 2-2) to 2-4) until the termination condition is met:

[0026]

[0027] Among them, σ is the discrimination accuracy and is greater than 0.

[0028] A further design of the regional renewable energy output and load forecasting method is that the coarser-scale nodes in the multi-resolution C-tree aggregate the C-node information of the corresponding finer-scale nodes.

[0029] A further design of the regional renewable energy output and load forecasting method is that the edges of the pyramid attention module can be divided into inter-scale connections and intra-scale connections, where the nodes of the finest scale correspond to the minimum time points of the time series, and the nodes of the coarser scale represent features of larger resolution.

[0030] The further design of the regional new energy output and load forecasting method is to optimize the model based on the set rules as follows:

[0031] Set s=1, 2, ..., S represents the sequence from the finest scale at the bottom to the coarse scale at the top in the pyramid attention module, represents the lth node at scale s; each node can follow a set of adjacent nodes at three scales Adjacent A nodes of the same scale, including the node itself, are represented as C child nodes are represented in a C-ary tree as and its parent node in the C-ary tree is represented as Then we have:

[0032]

[0033] node Attention y i The expression to be calculated is:

[0034]

[0035] Among them, L represents the sequence length, and A represents the number of adjacent nodes of the same scale; represents the number of attention layers of the lth node of scale s; d K Represents the query vector length, Q = XW Q ,q i ∈Q,K=XW K , k i ∈K,V=XW V , v l ∈V, X is the original input sample sequence, W Q , W K , W V is the weight matrix, Q, K, and V represent the query vector, key vector, and value vector in the attention mechanism, respectively. v l ∈V;

[0036] During the training process, the attention information expression of nodes at different time scales of the input sequence is optimized to obtain the focused feature information and improve the model's nonlinear fitting ability for sequence input and output at any time step.

[0037] A further design of the regional new energy output and load forecasting method is that in the step 3), the multimodal feature sequence, meteorological feature sequence, and time coding feature sequence are fused by random selection method or maximum entropy principle to obtain a multidimensional feature fusion sequence group.

[0038] The present invention also provides an electronic device, comprising a memory, a processor and a computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor to implement the regional new energy output and load forecasting method based on source-load common mode.

[0039] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the regional new energy output and load forecasting method based on source-load common mode.

[0040] The advantages of the present invention are as follows:

[0041] The regional renewable energy output and load forecasting method of the present invention jointly models the renewable energy output, electricity load data, meteorological data and other elements, so that the model can make full use of the temporal and spatial correlation information of the sequence, and enhance the common mode effect of the complex coupling relationship under the influence of multiple factors such as the supply side and the demand side.

[0042] This method decomposes the random fluctuation sequence into multiple regular modes in the frequency domain through variational mode decomposition operation, thereby improving the adaptive feature learning ability.

[0043] This method uses a pyramid C-tree structure to represent features, improves model learning efficiency in time and space dimensions, and reduces time complexity.

[0044] Through end-to-end model training, this method can learn sequence input and output of any time step, improve the model's long-term learning ability, and at the same time consider the data dependency between the source and load ends, better refer to different dimensions and more hidden information, and enhance the model's prediction ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of the method for regional renewable energy output and load forecasting based on source-load common mode of the present invention. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below.

[0047] like Figure 1 The method for regional new energy output and load forecasting based on source-load common mode of this embodiment includes the following steps:

[0048] Step 1) Collect the historical sequence of the source and load end and the meteorological data of the same period.

[0049] Step 2) By using the original sequence of variational mode decomposition, the multi-modal characteristic sequences of power load and renewable energy output sequence at different frequencies are obtained.

[0050] Step 3) Decompose the power load and new energy output sequences using different time scale features to obtain feature sequences of different period lengths in the time dimension; calculate the meteorological data of the corresponding area and the corresponding load sequence and new energy output sequence using the Pearson correlation coefficient method, select meteorological features with larger correlation coefficients, and combine them into meteorological feature sequences; perform one-hot encoding on the time label of the time dimension to obtain a time coding feature sequence; fuse the multimodal feature sequence, meteorological feature sequence, and time coding feature sequence to obtain a multidimensional feature fusion sequence group.

[0051] Step 4) constructing a model containing a pyramid attention module based on a multi-tree structure, and optimizing the model based on set rules; the pyramid attention module constructs a multi-resolution representation of the original sequence at different time scales, and a C-tree of a pyramid structure is formed by nodes of different time scales.

[0052] Step 5) Collect the features given by the last node of all scales in the pyramid attention module, map them to all future time steps in the batch, obtain the final prediction through a fully connected layer across the channel dimension, and output all future prediction results linearly together.

[0053] The model of this embodiment also includes: a data embedding module and a multi-time resolution C-tree module. The data embedding module embeds the acquired information including observation data, covariate data and location respectively, and then adds the various parameters in the information together to form a new feature representation. The multi-time resolution C-tree module (CSCM module) uses a coarse scale to construct a multi-resolution C-tree and completes information transfer between different scales through nodes. The model introduces a pyramid attention module (PAM module) so that the inter-scale tree structure summarizes the features at different resolutions, and the adjacent connections within the scale model the temporal dependencies of different ranges.

[0054] In step 2), variational mode decomposition (VMD) is a signal decomposition method that estimates each signal component by solving the frequency domain variational optimization problem, and decomposes the original sequence into a specified number of modal components from the frequency domain dimension; since the collected new energy output sequence and power load sequence have fluctuation signals of different frequencies, through disassembly and combination, the model can more easily capture the joint change law of the sequence in macro trends and micro mutations. VMD assumes that all components are narrowband signals concentrated near their respective center frequencies. Therefore, VMD establishes a constrained optimization problem based on the narrowband conditions of the components, thereby estimating the center frequency of the signal component and reconstructing the corresponding component. The decomposition process of VMD is the process of solving the variational problem. The algorithm mainly includes the construction of the variational problem and the solution of the variational problem.

[0055] First, define the intrinsic mode component (IMF), and then construct a variational problem, requiring that the sum of the bandwidths of the center frequency of each modal component is minimized, and the sum of all modal components is equal to the original signal. For each modal function, the unilateral spectrum is calculated through Hilbert transform, and then multiplied by the estimated center frequency, the unilateral spectrum of each mode is moved to the estimated center frequency, and then the bandwidth of each modal signal is calculated by the square norm of the gradient. Finally, by introducing the Lagrange operator and the quadratic penalty factor, the inequality constraint is transformed into an equality constraint for solution. The specific steps include the following:

[0056] Step 2-1) Initialize the modal function set Center frequency set Lagrangian operator λ 1 , initial iteration number b = 0;

[0057] Step 2-2) Set the number of iterations b = b + 1;

[0058] Step 2-3) When w ≥ 0 is satisfied, update the spectrum of each mode

[0059]

[0060] Update center frequency w k :

[0061]

[0062] Step 2-4) Update the Lagrange multiplier λ:

[0063]

[0064] Step 2-5) Repeat the above steps 2-2) to 2-4) until the termination condition is met:

[0065]

[0066] Among them, σ is the discrimination accuracy and is greater than 0.

[0067] A further design of the regional renewable energy output and load forecasting method is that the coarser-scale nodes in the multi-resolution C-tree aggregate the C-node information of the corresponding finer-scale nodes.

[0068] A further design of the regional renewable energy output and load forecasting method is that the edges of the pyramid attention module can be divided into inter-scale connections and intra-scale connections, where the nodes of the finest scale correspond to the minimum time points of the time series, and the nodes of the coarser scale represent features of larger resolution.

[0069] The further design of the regional new energy output and load forecasting method is to optimize the model based on the set rules as follows:

[0070] Set s=1, 2, ..., S represents the finest scale at the bottom to the coarse scale at the top of the pyramid attention module, represents the lth node at scale s; each node can follow a set of adjacent nodes at three scales Adjacent A nodes of the same scale, including the node itself, are represented as C child nodes are represented in a C-ary tree as and its parent node in the C-ary tree is represented as Then we have:

[0071]

[0072] point Attention y i The expression to be calculated is:

[0073]

[0074] N represents the number of attention layers.

[0075] VMD is a non-recursive, adaptive signal decomposition method that can decompose non-stationary, nonlinear complex signals and effectively suppress modal aliasing. The IMFs decomposed by the VMD method in this embodiment all have independent center frequencies and exhibit sparse characteristics in the frequency domain.

[0076] In step 3), the multimodal feature sequence, the meteorological feature sequence, and the time coding feature sequence are fused by random selection method or maximum entropy principle to obtain a multidimensional feature fusion sequence group.

[0077] This embodiment also provides an electronic device, including a memory, a processor, and a computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor to implement the regional new energy output and load forecasting method based on source-load common mode.

[0078] This embodiment also provides a computer-readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the regional new energy output and load forecasting method based on source-load common mode.

[0079] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for forecasting regional renewable energy output and load based on source-load common mode, characterized in that The steps include: Step 1) Collect the historical sequence of the source and load end and the meteorological data of the same period, Step 2) obtaining the multi-modal characteristic sequences of the power load and the new energy output sequence at different frequencies through the original sequence of variational mode decomposition; Step 3) Decomposing the power load and new energy output sequences by using different time scale features to obtain feature sequences of different period lengths in the time dimension; calculating the meteorological data of the corresponding area and the corresponding load sequence and the new energy output sequence by using the Pearson correlation coefficient method, selecting meteorological features with larger correlation coefficients, and combining them into meteorological feature sequences; performing unique hot encoding on the time labels of the time dimension to obtain a time coding feature sequence; fusing the multimodal feature sequence, the meteorological feature sequence, and the time coding feature sequence to obtain a multidimensional feature fusion sequence group; Step 4) constructing a model containing a pyramid attention module based on a multi-tree structure, and optimizing the model based on set rules; the pyramid attention module constructs a multi-resolution representation of the original sequence at different time scales, and the C-tree of the pyramid structure is composed of nodes of different time scales; Step 5) Collect the features given by the last node of all scales in the pyramid attention module, map them to all future time steps in the batch, obtain the final prediction through a fully connected layer across the channel dimension, and output all future prediction results linearly together.

2. The regional new energy output and load forecasting method according to claim 1 is characterized in that The model also includes: The data embedding module constructs a main input variable containing information of observation data, covariate data and position through modal decomposition and multi-feature selection according to importance, embeds the main input variable, and combines various parameters in the information by constructing a tree-structured attention mechanism to form a new high-dimensional feature representation; The multi-time resolution C-tree module uses a coarse scale to construct a multi-resolution C-tree and completes the information transfer between different scales through nodes.

3. The regional new energy output and load forecasting method according to claim 1 is characterized in that The variational mode decomposition in step 2) specifically includes the following steps: Step 2-1) Initialize the modal function set Center frequency set Lagrangian operator λ 1 , initial iteration number b = 0; Step 2-2) Set the number of iterations b = b + 1; Step 2-3) When w ≥ 0, update the spectrum of each mode Among them, α is the penalty coefficient, which determines the bandwidth of the component mode function; w is the center frequency, w k represents the center frequency of the kth mode; The Lagrangian operator representing the signal to be decomposed; Represents the frequency domain representation of the signal to be decomposed Update center frequency w k : Step 2-4) Update the Lagrange multiplier λ: Among them, ε is the noise tolerance parameter; Step 2-5) Repeat the above steps 2-2) to 2-4) until the termination condition is met: Among them, σ is the discrimination accuracy and is greater than 0.

4. The method for forecasting regional renewable energy output and load according to claim 3 is characterized in that The coarser-scale nodes in the multi-resolution C-tree summarize the C-node information of the corresponding finer-scale nodes.

5. The method for forecasting regional renewable energy output and load according to claim 2, characterized in that The edges of the pyramid attention module can be divided into inter-scale connections and intra-scale connections, where the nodes at the finest scale correspond to the smallest time point of the time series, and the nodes at the coarser scale represent features with larger resolution.

6. The method for forecasting regional renewable energy output and load according to claim 2, characterized in that The optimization of the model based on the set rules is specifically as follows: Set s=1, 2, ..., S represents the sequence from the finest scale at the bottom to the coarse scale at the top in the pyramid attention module, represents the lth node at scale s; each node can follow a set of adjacent nodes at three scales Adjacent A nodes of the same scale, including the node itself, are represented as C child nodes are represented in a C-ary tree as and its parent node in the C-ary tree is represented as Then we have: node Attention y i The expression to be calculated is: Among them, L represents the sequence length, and A represents the number of adjacent nodes of the same scale; represents the number of attention layers of the lth node of scale s; dK represents the query vector length, Q = XW Q ,q i ∈Q,K=XW K , k i ∈K,V=XW V , v l ∈V, X is the original input sample sequence, W Q , W K , W V is the weight matrix, Q, K, and V represent the query vector, key vector, and value vector in the attention mechanism, respectively. v l ∈V; During the training process, the attention information expression of nodes at different time scales of the input sequence is optimized to obtain the focused feature information and improve the model's nonlinear fitting ability for sequence input and output at any time step.

7. The regional renewable energy output and load forecasting method according to claim 1 is characterized in that In the step 3), the multimodal feature sequence, the meteorological feature sequence, and the time coding feature sequence are fused by random selection method or maximum entropy principle to obtain a multidimensional feature fusion sequence group.

8. An electronic device, characterized in that: It includes a memory, a processor and a computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor to implement the regional new energy output and load forecasting method based on source-load common mode as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and the computer program is executed by a processor to implement the regional new energy output and load forecasting method based on source-load common mode as described in any one of claims 1-7.

Images