CN-121981314-A - Agricultural load prediction method and system based on multi-component time sequence decoupling multi-mode learning

CN121981314ACN 121981314 ACN121981314 ACN 121981314ACN-121981314-A

Abstract

The invention discloses an agricultural load prediction method and system based on multi-mode learning with multi-component time sequence decoupling, and belongs to the technical field of agricultural load prediction. The method comprises the steps of collecting historical agricultural load and meteorological data, decomposing the historical agricultural load and the meteorological data by utilizing a multivariate variation mode to obtain periodic, trend and residual modal components, respectively constructing a time convolution neural network, a bidirectional gating circulation unit and a support vector regression model aiming at decomposed periodic, trend and residual modal component data sets, and fully mining characteristic information of each mode after decomposition, so that accurate prediction of the agricultural load is realized. The method captures the potential nonlinear space-time coupling relation between the agricultural load and the meteorological factors, improves the prediction effect of the agricultural load under the seasonal period fluctuation scene, the long-term trend and the abnormal scene of the agricultural load, improves the prediction precision of the agricultural load, and provides support for the reliable and stable operation of the power grid.

Inventors

MAO XIAOBO
YONG YE
Xue Mifeng

Assignees

国网江苏省电力有限公司无锡供电分公司

Dates

Publication Date: 20260505
Application Date: 20251218

Claims (10)

1. The agricultural load prediction method based on multi-modal learning of multi-component time sequence decoupling is characterized by comprising the following steps: collecting historical agricultural load and meteorological data; performing multi-element time sequence decoupling on historical agricultural loads and meteorological data by utilizing MVMD to obtain periodic, trend and residual modal components; Constructing a TCN model, a Bi-GRU model and an SVR model; Respectively inputting the period, the trend and the residual modal components into a constructed TCN model, a Bi-GRU model and a SVR model for learning, and respectively excavating agricultural load prediction information corresponding to the period, the trend and the residual modal components; and combining the excavated agricultural load prediction information corresponding to the period, trend and residual modal components to obtain an agricultural load prediction result, and realizing the agricultural load prediction based on multi-component time sequence decoupling multi-modal learning.
2. The agricultural load prediction method based on multi-sequential decoupling multi-modal learning according to claim 1, wherein: The historical agricultural load and meteorological data comprises agricultural load, temperature, humidity, wind speed and precipitation.
3. The agricultural load prediction method based on multi-sequential decoupling multi-modal learning according to claim 1, wherein: the method for performing multi-element time sequence decoupling on historical agricultural loads and meteorological data by utilizing MVMD to obtain periodic, trend and residual modal components specifically comprises the following steps: constructing a multichannel signal matrix according to historical agricultural loads and meteorological data; constructing a mode matrix to be solved according to the multi-channel signal matrix; constructing an agricultural data decomposition model according to a mode matrix to be solved by taking the minimum total spectrum bandwidth as a target; constructing a Lagrangian function, and converting an agricultural data decomposition model into an unconstrained optimization problem; and solving an unconstrained optimization problem by using an alternate direction multiplier method, wherein signals of each channel are decoupled into 3 modal components, and periodic, trend and residual modal components are obtained.
4. The agricultural load prediction method based on multi-modal learning with multi-component time sequence decoupling as claimed in claim 3, wherein: the multi-channel signal matrix is constructed according to the historical agricultural load and the meteorological data and is expressed by the following formula: wherein: In the case of a multi-channel signal, Is that Is the first of (2) Each component, C is The total number of components, In order to be used as a load for agriculture, In order to be able to determine the temperature, In order to be a degree of humidity, For the wind speed of the wind, Is precipitation.
5. The agricultural load prediction method based on multi-modal learning with multi-component time sequence decoupling as claimed in claim 3, wherein: The mode matrix to be solved is constructed according to the multi-channel signal matrix and is expressed by the following formula: wherein: In the case of a multi-channel signal, Is that Is the first of (2) Each component, C is The total number of components, Representation of Is selected from the group consisting of the (c) and (d), To set the number of modes to be decomposed, let 。
6. The agricultural load prediction method based on multi-modal learning with multi-component time sequence decoupling as claimed in claim 3, wherein: the method aims at minimizing the total spectrum bandwidth, constructs an agricultural data decomposition model according to a mode matrix to be solved, and is expressed by the following formula: wherein: Is that Is the first of (2) The number of components of the composition, Is a multi-channel signal, C is The total number of components, Indicating that the time t is biased, Representation of Is selected from the group consisting of the (c) and (d), Is the center frequency of the kth mode after MVMD decomposition, Is that Is used for the analysis of the modulation function, To set the number of modes to be decomposed, let 。
7. The agricultural load prediction method based on multi-sequential decoupling multi-modal learning according to claim 1, wherein: The periodic, trend and residual modal components are respectively input into a constructed TCN model, a Bi-GRU model and an SVR model for learning, and the periodic modal components are input into the TCN model, which comprises the following steps: Compressing and linearly transforming the characteristic dimension of the periodic modal component through a1 multiplied by 1 convolution layer to generate a one-dimensional agricultural load characteristic retaining the original periodic characteristic; Setting an expansion coefficient of a periodic modal component through first expansion causal convolution, capturing an agricultural load characteristic of an inherent time scale of the periodic modal component, and sequentially carrying out first weight normalization, first ReLU activation and first Dropout regularization on the captured characteristic to obtain the agricultural load characteristic of the inherent time scale; capturing the time sequence dependent agricultural load characteristics of the periodic modal components through second dilation causal convolution by the inherent time scale agricultural load characteristics, and sequentially carrying out second weight normalization, second ReLU activation and second Dropout regularization on the captured characteristics to obtain regularized agricultural load characteristics; and carrying out residual connection and element-by-element addition on the one-dimensional agricultural load characteristic which keeps the original periodic characteristic and the regularized agricultural load characteristic to generate an agricultural load predicted value of the periodic modal component.
8. The agricultural load prediction method based on multi-sequential decoupling multi-modal learning according to claim 1, wherein: The periodic, trend and residual modal components are respectively input into a constructed TCN model, a Bi-GRU model and a SVR model for learning, and the trend modal components are input into the Bi-GRU model, which comprises the following steps: Inputting the trend modal components into a forward GRU according to forward time sequence from the past to the future, and capturing the historical evolution rule characteristics of the agricultural load; inputting trend modal components into a reverse GRU according to a reverse time sequence from the past to the future, and capturing future change direction characteristics of agricultural loads; and adding the historical evolution rule characteristic of the agricultural load and the future change direction characteristic of the agricultural load, inputting the characteristics into the full-connection layer, and generating an agricultural load predicted value of the trend modal component.
9. The agricultural load prediction method based on multi-sequential decoupling multi-modal learning according to claim 1, wherein: the period, trend and residual modal components are respectively input into a constructed TCN model, a Bi-GRU model and an SVR model for learning, and the residual modal components are input into the SVR model, which comprises the following steps: inputting the residual modal component into a Gaussian kernel function, and mapping the nonlinear characteristic of the residual modal component into a high-dimensional characteristic space; Constructing a convex optimization model by taking the problem of minimizing convex optimization as a target; solving the convex optimization model by using the sequence minimization to obtain an optimal weight vector and bias; inputting the optimal weight vector and bias into a high-dimensional feature space to construct a residual agricultural load prediction model; and inputting the nonlinear characteristic of the residual modal component mapped to the high-dimensional characteristic space into a residual agricultural load prediction model to obtain an agricultural load prediction value of the residual modal component.
10. An agricultural load prediction system based on multi-sequential decoupling multi-modal learning, operating the agricultural load prediction method based on multi-sequential decoupling multi-modal learning of any one of claims 1 to 9, characterized in that: The data acquisition module is used for acquiring historical agricultural loads and meteorological data; The modal decomposition module is used for performing multi-element time sequence decoupling on the historical agricultural load and the meteorological data by utilizing MVMD to obtain periodic, trend and residual modal components; the model building module is used for building a TCN model, a Bi-GRU model and an SVR model; The agricultural load prediction module is used for respectively inputting the periodic, trend and residual modal components into the constructed TCN model, bi-GRU model and SVR model for learning, and respectively excavating agricultural load prediction information corresponding to the periodic, trend and residual modal components; And the output module is used for merging the excavated agricultural load prediction information corresponding to the period, the trend and the residual modal component to obtain an agricultural load prediction result, and realizing the agricultural load prediction based on multi-modal learning by multi-element time sequence decoupling.

Description

Agricultural load prediction method and system based on multi-component time sequence decoupling multi-mode learning Technical Field The invention belongs to the technical field of agricultural load prediction, and particularly relates to an agricultural load prediction method based on multi-component time sequence decoupling multi-mode learning. Background In agricultural production, from irrigation systems, agricultural machinery to agricultural product processing and storage, power has become an indispensable production element, highly dependent on stable and reliable power supply. According to the data issued by the national energy agency, the cumulative electricity consumption of the agriculture in 2023 years is 1278 hundred million kilowatt-hours, the same ratio is increased by 11.5%, and the agricultural load demands are growing. Therefore, in order to ensure agricultural power supply, accurate agricultural load prediction is particularly important. However, agricultural load prediction is faced with unique challenges, agricultural electricity is affected by more external and uncontrollable factors, such as climate conditions, rainfall, crop type, planting area, agricultural machinery and skill level, groundwater level, etc., existing agricultural climate data are not only interrelated, but also have a high degree of regional and time variability impact, agricultural loads also have obvious seasonal characteristics, such as periodic demand for electricity demand during sowing, growing and harvesting periods, long-term climate change, and different agricultural electricity demand during extreme weather events. The prior art aims at the data of different time scales such as periodic variation of the agricultural power consumption according to season precipitation and the like, abrupt change of the agricultural power consumption according to the trend of the agricultural power consumption caused by long-term climate variation and extreme weather events, and the like, can only be used for processing single variables (such as only agricultural load or single meteorological factor), cannot jointly decompose the agricultural load and the meteorological factor, is difficult to capture the potential nonlinear space-time coupling relation between the agricultural load and the meteorological factor, displays separation period, trend and residual error components, causes the actual disconnection of the agricultural load plan, for example, does not recognize the coupling relation between precipitation and the agricultural load of irrigation equipment, can excessively depend on electric irrigation in drought period, causes the overload of an agricultural power grid, does not recognize the coupling relation between strong precipitation and the agricultural load surge power consumption of irrigation equipment, causes the early warning signal of lacking residual error components in the agricultural load prediction, and causes the abnormal power grid scheduling. In the agricultural load prediction method in the prior art, only a load sequence is decomposed, complex nonlinear space-time coupling relations among various input data cannot be fully considered, the same model is adopted for different modal components obtained through decomposition, the difference of different time scale characteristics is ignored, and the same model is used for agricultural data of different time scales, so that the agricultural load prediction is inaccurate. Disclosure of Invention In order to solve the defects in the prior art, the invention provides an agricultural load prediction method based on multi-sequential decoupling multi-mode learning, which utilizes MVMD (Multivariate Variational Mode Decomposition, multi-variable variation modal decomposition) to perform multi-sequential decoupling on historical agricultural load and meteorological data, respectively inputs periodic, trend and residual modal components into a constructed TCN (Temporal Convolutional Network, time convolution network) model, a Bi-GRU (Bidirectional Gated Recurrent Unit, a Bi-directional gating circulating unit) model and a SVR (Support Vector Regression ) model for learning, respectively excavates agricultural load prediction information corresponding to periodic, trend and residual modal components, and improves agricultural load prediction precision. The invention adopts the following technical scheme. The first aspect of the invention provides an agricultural load prediction method based on multi-modal learning by multi-component time sequence decoupling. Collecting historical agricultural load and meteorological data; performing multi-element time sequence decoupling on historical agricultural loads and meteorological data by utilizing MVMD to obtain periodic, trend and residual modal components; Constructing a TCN model, a Bi-GRU model and an SVR model; Respectively inputting the period, the trend and the residual modal components into a constructed TCN mo