CN-122026457-A - Distributed energy storage regulation and control method and system based on lightweight neural network

CN122026457ACN 122026457 ACN122026457 ACN 122026457ACN-122026457-A

Abstract

The invention discloses a distributed energy storage regulation and control method and a system based on a lightweight neural network, wherein the method comprises the steps of collecting operation data of distributed energy storage nodes and constructing a time sequence input characteristic matrix; the method comprises the steps of taking a time sequence input feature matrix as input of a lightweight one-dimensional convolution network to obtain local time feature vectors, taking the local time feature vectors as input of a lightweight gating circulation unit to obtain a time sequence feature matrix, and carrying out power and feasible domain projection after feature fusion of the local time feature vectors and the time sequence feature matrix. The invention can ensure the regulation and control precision, reduce the calculation cost and improve the autonomy, the instantaneity and the operation stability of the distributed energy storage system.

Inventors

MA XIAOBING
LIU XIAOYE
YUAN YU
GUO JINXING
XU BIN
MA TAO
WANG DAOJING
GUAN YAFEI
ZHAO YONGYI
GAO MAOCHENG
MA WEI
CAO YONG

Assignees

国网安徽省电力有限公司淮北供电公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. A distributed energy storage regulation and control method based on a lightweight neural network is characterized by comprising the following steps: collecting operation data of distributed energy storage nodes, and constructing a time sequence input feature matrix; taking the time sequence input feature matrix as the input of a lightweight one-dimensional convolution network to obtain a local time feature vector; Taking the local time feature vector as the input of a lightweight gating circulation unit to acquire a time sequence feature matrix; And after the local time feature vector and the time sequence feature matrix are fused, carrying out power and feasible domain projection.
2. The distributed energy storage regulation and control method based on the lightweight neural network is characterized in that the network architecture of the lightweight one-dimensional convolutional network is that the quantity of convolution kernels is adjusted from standard convolution to low, sparse and quantization constraints are introduced into a learnable weight of a convolutional layer, fixed-point quantization is introduced into the learnable weight of the convolutional layer, storage and calculation overhead is reduced, a channel gating compression mechanism is introduced into a convolutional channel, and the quantity of the convolutional channel is dynamically adjusted.
3. The method for controlling distributed energy storage based on a lightweight neural network according to claim 2, wherein the introducing of sparsity and quantization constraints to the learnable weights of the convolutional layers comprises: Rearranging the convolution kernel weights into a forming matrix to be used as a convolution layer weight parameter set, applying sparse regularization constraint to the convolution layer weight parameter set, and setting the corresponding weights to be zero if the weight elements in the convolution layer weight parameter set are smaller than a pruning threshold value.
4. The method for regulating and controlling distributed energy storage based on the lightweight neural network according to claim 2, wherein introducing fixed-point quantization to the learnable weights of the convolution layer reduces storage and calculation overhead, and comprises performing 8-bit fixed-point quantization to the convolution kernel weights, wherein all floating point weights are mapped into integer weights in the quantization process.
5. The distributed energy storage regulation and control method based on the lightweight neural network according to claim 2 is characterized in that a channel gating compression mechanism is introduced into a convolution channel, the number of the convolution channels is dynamically regulated, the method comprises the steps of introducing a channel gating function into the convolution channel, setting a gating pruning threshold value, and if the channel gating function is smaller than the gating pruning threshold value, the lightweight one-dimensional convolution network prunes convolution calculation corresponding to the convolution channel in an inference stage.
6. The distributed energy storage regulation and control method based on the lightweight neural network, which is characterized in that the network architecture of the lightweight gating circulation unit is that low-rank decomposition and parameter upper limit constraint are applied to parameters of the lightweight gating circulation unit, learning gating is introduced to hidden units to perform gating sparsity and neuron clipping, and fixed-point quantization is introduced to a parameter matrix optimized by low-rank decomposition to reduce storage and calculation cost.
7. The method for controlling distributed energy storage based on a lightweight neural network according to claim 6, wherein applying low-rank decomposition and parameter upper limit constraint to the parameters of the lightweight gating cycle unit comprises: and carrying out low-rank decomposition on the input weight and the circulating weight of the lightweight gating circulating unit: ; ; Wherein, the In order to input the weight(s), A parameter matrix optimized for low rank decomposition corresponding to the input weights, For the transposition of the matrix, Is a real number, and is a real number, For the hidden state dimension of the lightweight gated loop unit, For a low-rank decomposed rank, ; As the dimension of the local temporal feature vector, The representation takes on a smaller value for both, For the cyclic weight of the signal, the signal is weighted, A parameter matrix optimized for low-rank decomposition corresponding to the cyclic weights; The parameter amount of the lightweight gating loop unit is reduced to: ; In the formula, The reduced parameter number is the lightweight gating cycle unit.
8. The method for controlling distributed energy storage based on a lightweight neural network according to claim 1, wherein performing power and feasible-area projection comprises: executing the rated power and single-step change rate constraint of the equipment on the fused features to obtain an active power instruction after projection, and obtaining a power projection result which simultaneously meets the limit of the power amplitude and the climbing rate; based on the power projection result, carrying out state of charge consistency constraint projection; and integrating the consistency constraint of the equipment and the charge state to obtain the executable target charge and discharge power.
9. The method for regulating and controlling distributed energy storage based on a lightweight neural network according to claim 8, wherein the target charge-discharge power is expressed by the following formula: ; In the formula, For the purpose of the charge-discharge power, Is the first At each energy storage node The projection result of the limit of the power amplitude and the climbing rate is simultaneously satisfied at the moment, 、 Are respectively the first The minimum/maximum adjustable active power limit of the individual energy storage nodes, 、 Respectively is the front part Under the condition of meeting the constraint of state of charge, the minimum/maximum active power allowed is a function , The minimum allowable value and the maximum allowable value are respectively the input quantity to be limited.
10. A system for applying the lightweight neural network-based distributed energy storage regulation method of any one of claims 1-9, comprising: the acquisition module is used for acquiring the operation data of the distributed energy storage nodes and constructing a time sequence input characteristic matrix; The convolution network module is used for taking the time sequence input feature matrix as the input of the lightweight one-dimensional convolution network to acquire a local time feature vector; The gating circulation module is used for taking the local time feature vector as the input of the lightweight gating circulation unit to acquire a time sequence feature matrix; and the feasible domain projection module is used for carrying out power and feasible domain projection after the local time feature vector and the time sequence feature matrix feature are fused.

Description

Distributed energy storage regulation and control method and system based on lightweight neural network Technical Field The invention relates to the technical field of intelligent regulation and control of a distributed energy storage system and energy Internet, in particular to a distributed energy storage regulation and control method and system based on a lightweight neural network. Background Along with the high-proportion access of renewable energy sources such as distributed photovoltaic, wind power and the like, the power fluctuation and uncertainty of a power distribution side are obviously increased, and distributed energy storage (BESS) is widely used for peak clipping and valley filling, frequency modulation support and accident standby. The existing energy storage regulation and control multi-dependency centralized optimization or complex depth model is characterized in that the former needs to frequently send full data, communication delay and single-point fault risk are high, the latter has large parameter quantity and high calculation power and energy consumption cost, and the former is difficult to stably run on an edge controller (embedded MCU/ARM) in real time. Meanwhile, the battery SOC/SOH, temperature and power constraint have strong coupling and time variability, the traditional threshold value/rule strategy is difficult to consider the safety boundary and economy, and the problems of response lag, out-of-range charge and discharge, frequent power jitter and the like are easy to occur. Therefore, an intelligent regulation and control method which is oriented to the edge side, low in parameter quantity and low in time delay and can be cooperated with physical constraint is urgently needed, on-site feature extraction, time sequence modeling and feasible domain projection are realized, communication dependence is reduced, and system robustness and instantaneity are improved. Based on the method, a distributed energy storage regulation method fused with a gating circulation unit by utilizing light one-dimensional convolution under an edge computing environment is provided. The invention patent with the publication number of CN119026888A discloses a method and a system for regulating and controlling the sharing energy storage of a platform region based on the prediction of the demand of the sharing energy storage, wherein the centralized sharing energy storage of the platform region level is emphasized, the demand prediction is realized by combining a CNN-GRU-Attention combined prediction model through screening on similar days, a prediction result is used as the input of multi-objective optimization regulation and control, and finally the optimal regulation and control strategy of the platform region is solved. Namely, the patent is directed to a centralized or shared energy storage system of a platform, a regulation object is a unified energy storage resource of the platform layer, the combination of CNN-GRU is taken as a prediction model, the output of the prediction model is used for a subsequent optimization algorithm, and the prediction model is not specially designed for edge deployment, embedded calculation force constraint or real-time reasoning. Disclosure of Invention The invention aims to solve the technical problems that the centralized computation time delay is large, the model is complex and difficult to deploy at the edge end, the multi-constraint coordination on SOC/temperature/power and the like is insufficient, the communication bandwidth occupation is high and the like in the existing distributed energy storage regulation and control. In order to solve the technical problems, the invention provides the following technical scheme: a distributed energy storage regulation and control method based on a lightweight neural network comprises the following steps: collecting operation data of distributed energy storage nodes, and constructing a time sequence input feature matrix; taking the time sequence input feature matrix as the input of a lightweight one-dimensional convolution network to obtain a local time feature vector; Taking the local time feature vector as the input of a lightweight gating circulation unit to acquire a time sequence feature matrix; And after the local time feature vector and the time sequence feature matrix are fused, carrying out power and feasible domain projection. In the embodiment, the network architecture of the lightweight one-dimensional convolution network is that the number of convolution kernels is adjusted from standard convolution to low, sparse and quantization constraints are introduced into a learnable weight of a convolution layer, fixed-point quantization is introduced into the learnable weight of the convolution layer, storage and calculation overhead is reduced, a channel gating compression mechanism is introduced into a convolution channel, and the number of the convolution channels is dynamically adjusted. In this embodiment, intr