Search

CN-122026317-A - Distributed photovoltaic short-term power prediction method and system based on improved clustering and parallel feature extraction

CN122026317ACN 122026317 ACN122026317 ACN 122026317ACN-122026317-A

Abstract

A distributed photovoltaic short-term power prediction method based on improved clustering and parallel feature extraction aims to solve the problem of insufficient prediction precision caused by large distributed photovoltaic quantity, wide regional distribution and large output fluctuation. The method comprises the steps of constructing an improved Deep Convolution Embedded Clustering (DCEC) algorithm, embedding a time sequence-space dual-attention module in a DCEC encoder, realizing collaborative extraction of data driving features and space-time features, improving intra-cluster compactness and inter-cluster separability of clusters, constructing a CNN-BiGRU parallel prediction framework, extracting inter-cluster coupling features and weather-photovoltaic correlation features through CNN, capturing photovoltaic power bidirectional time sequence dependent features by BiGRU, splicing the photovoltaic power bidirectional time sequence dependent features through a fusion layer to form a high-dimensional prediction representation, and outputting a short-term photovoltaic power prediction result through a full-connection layer and an output layer. Compared with the traditional method, the method has the advantages that the prediction error is obviously reduced, the prediction stability and generalization capability under the complex weather condition are improved, and the method can effectively support the application of power grid dispatching, energy storage optimization and the like.

Inventors

  • ZHANG YUBAO
  • LI MENG
  • LI YAOKUN
  • WANG CHUNPENG
  • LI FEIFEI

Assignees

  • 国网山东省电力公司德州供电公司

Dates

Publication Date
20260512
Application Date
20260104

Claims (11)

  1. 1. The distributed photovoltaic short-term power prediction method based on improved clustering and parallel feature extraction is characterized by comprising the following steps of: step 1, data preprocessing, namely acquiring historical power data of a distributed photovoltaic power station and corresponding multidimensional meteorological data, carrying out normalization processing, and constructing a power matrix data set taking date-time as a two-dimensional structure; Step 2, deep feature learning and clustering based on a time sequence-space dual-attention mechanism: Step 2.1, an improved depth convolution embedded clustering model is constructed, a time sequence attention sub-module and a space attention sub-module are sequentially embedded in an encoder, and the time sequence attention sub-module and the space attention sub-module are used for cooperatively extracting time sequence dependent features and space associated features from the power matrix data set to generate space-time fusion enhancement features; Step 2.2, taking the space-time fusion enhancement characteristic as input, carrying out soft distribution through a clustering layer, introducing a clustering loss function of KL divergence and a reconstruction loss joint loss function of an encoder-decoder, and training the improved depth convolution embedded clustering model; step 2.3, clustering the power matrix data set by using the improved deep convolution embedded clustering model after training to obtain a plurality of clustering clusters with different photovoltaic output physical modes, and extracting characteristic data corresponding to the corresponding clustering clusters; Step3, feature extraction and prediction based on parallel architecture: step 3.1, constructing a CNN-BiGRU parallel prediction framework which comprises a CNN branch and a BiGRU branch which are arranged in parallel, and a fusion layer, a full connection layer and an output layer; step 3.2, inputting the cluster feature data output in the step 2 and multidimensional meteorological data into the CNN-BiGRU parallel prediction architecture, and simultaneously feeding a CNN branch and a BiGRU branch of the parallel prediction architecture, wherein the CNN branch is used for extracting inter-cluster coupling features contained in the cluster feature data and nonlinear correlation features between the cluster feature data and the meteorological data, and the BiGRU branch is used for extracting forward and backward bidirectional time sequence dependence features from the cluster feature data; step 3.3, splicing the output characteristics of the CNN branch and the BiGRU branch at the fusion layer to form a high-dimensional joint characteristic vector; Step 3.4, the high-dimensional joint feature vector is subjected to nonlinear mapping through the full-connection layer and regression through the output layer in sequence to generate a short-term photovoltaic power predicted value; And 4, model optimization and verification, namely performing end-to-end or staged joint training and optimization on the parameters of the improved deep convolution embedded clustering algorithm model in the step 2 and the CNN-BiGRU parallel prediction framework in the step 3 by adopting an optimization algorithm, and inhibiting model overfitting by adopting a regularization strategy to obtain a final distributed photovoltaic short-term power prediction model.
  2. 2. The method according to claim 1, wherein the implementation of the time sequence attention sub-module in step 2.1 is as follows: taking the date-time power matrix as input, adopting one-dimensional convolution operation to extract power dependency characteristics of adjacent time along a time axis in a sliding way, and obtaining a time sequence characteristic diagram; global average pooling is carried out on the time sequence feature diagram along a date axis and a channel axis to obtain a time sequence feature compression vector; generating attention weight of time dimension by using the time sequence feature compression vector through a Sigmoid activation function; And multiplying the time sequence characteristic diagram by the attention weight element by element to obtain the time sequence enhancement characteristic.
  3. 3. The method for predicting the short-term power of a distributed photovoltaic according to claim 2, wherein the implementation process of the spatial attention sub-module in step 2.1 is as follows: taking the time sequence enhancement feature as input, extracting space correlation features crossing time-date by adopting two-dimensional convolution operation, and obtaining a space feature map; global average pooling is carried out on the space feature map to obtain a space feature compression vector; generating the attention weight of the date dimension by the space feature compression vector through a network comprising two full-connection layers and a Sigmoid activation function; multiplying the space feature map with the attention weight element by element to obtain a space-time fusion enhancement feature; and the space-time fusion enhancement features are processed by a pooling layer and a convolution layer, and then low-dimensional feature vectors are obtained through global average pooling operation and are used for subsequent clustering.
  4. 4. The method according to claim 1, wherein the step 2.2 of improving the DCEC algorithm has a clustering loss function of KL divergence for minimizing the difference between the soft allocation probability and the target distribution obtained by squaring the soft allocation probability to enhance the weight of the high confidence samples.
  5. 5. The method according to claim 1, wherein the structure of the CNN module in step 3 includes a convolution layer and a pooling layer, the convolution layer uses a ReLU activation function, the pooling layer uses a maximum pooling operation, and local pattern features are extracted through a local connection and weight sharing mechanism, so as to reduce data dimensions.
  6. 6. The method according to claim 1, wherein the BiGRU modules in step 3 include a forward GRU and a backward GRU, the forward GRU models time-series dependence along a time axis, the backward GRU models time-series dependence against the time axis, and hidden states of the two modules are spliced to be output features of the BiGRU module.
  7. 7. The method according to claim 1, wherein the feature fusion manner of the fusion layer in the step 3 is feature stitching, the output of the BiGRU module is h_t, the output of the CNN module is o_t, and the fusion feature t= Concat (h_t, o_t) = [ h_to_t ].
  8. 8. The method according to claim 1, wherein the regularization rate of the Dropout strategy in step 4 is set to 0.6, the learning rate of adam algorithm is set to 0.001, the number of model iterations is set to 1000, and the loss function uses Mean Square Error (MSE).
  9. 9. The distributed photovoltaic short-term power prediction method according to claim 1, wherein the multi-dimensional meteorological data comprises at least one of solar irradiance, temperature, humidity, rainfall.
  10. 10. The method according to claim 1, wherein the number of clusters in step 2 is determined by Dunn Index (DI) and Calinski-Harabasz index (CH), and the optimal number of clusters is a value that enables both indexes to reach local maximum.
  11. 11. A distributed photovoltaic short-term power prediction system based on improved clustering and parallel feature extraction, comprising: the data acquisition and preprocessing module is used for acquiring historical power data of the distributed photovoltaic power station and corresponding multidimensional meteorological data, and constructing a date-time power matrix data set after normalization processing is carried out on the acquired data; The improved clustering module is used for clustering the power matrix data set by adopting an improved DCEC algorithm embedded with a time sequence-space dual-attention module, wherein a time sequence attention sub-module and a space attention sub-module are built in the module, so that the collaborative extraction of time sequence dependence and space association characteristics is realized, and cluster feature data with a definite force output mode is output; the parallel feature extraction and prediction module is used for constructing a CNN-BiGRU parallel prediction framework, inputting the cluster feature data and the multidimensional meteorological data into the framework, extracting inter-cluster coupling features and meteorological-photovoltaic correlation features through a CNN submodule, extracting bidirectional time sequence dependence features through a BiGRU submodule, splicing and fusing the two types of features through a fusion submodule, and outputting a short-term photovoltaic power predicted value through a full-connection submodule and an output submodule; the model optimization module is used for optimizing parameters of the improved clustering module and the parallel feature extraction and prediction module by adopting an Adam algorithm, inhibiting overfitting by using a Dropout strategy and improving the generalization capability of the model; And the prediction result output module is used for formatting and outputting the short-term photovoltaic power prediction value and providing data support for power grid dispatching, energy storage optimization and other applications.

Description

Distributed photovoltaic short-term power prediction method and system based on improved clustering and parallel feature extraction Technical Field The invention belongs to the technical field of power system automation and intelligent prediction, and particularly relates to a distributed photovoltaic short-term power prediction method and system based on improved clustering and parallel feature extraction, which are particularly suitable for power system scenes with high-proportion distributed photovoltaic access and output form showing fragmentation and heterogeneity characteristics, and can effectively support related applications such as power grid dispatching, energy storage optimization, demand side management and the like. Background With the rapid deployment of distributed photovoltaics in town buildings, industrial and commercial parks and rural roofs, the access quantity of the distributed photovoltaics is exponentially increased, and the output form shows the remarkable characteristics of fragmentation, heterogeneity and diversification. A large number of small-scale photovoltaic units are distributed under different roof, orientation, tilt and shading conditions, so that the power output of the photovoltaic units has stronger uncertainty and fluctuation than centralized photovoltaic. Short-term power prediction is used as a key technology for supporting power grid dispatching, energy storage optimization, demand side management and market trading, and the accuracy of the short-term power prediction directly influences grid-connected friendliness and digestion capability of distributed photovoltaic. However, since distributed photovoltaics generally lack high-precision meteorological measurement equipment and have wide-point and multi-surface distribution characteristics, the performance of the existing prediction method still cannot meet engineering application requirements, and research on the high-precision prediction method facing the distributed scene is needed to be carried out. At present, research on the prediction of distributed photovoltaic power at home and abroad is still in a preliminary development stage, and a plurality of researches still use a centralized photovoltaic technical route, so that prediction errors are large. The existing research shows that the error of the distributed photovoltaic prediction is generally 3% -8% higher than that of the centralized photovoltaic prediction, and the complexity of the distributed scene is fully reflected in the difficulty of adapting to the traditional method. In order to improve the prediction performance of the distributed photovoltaic, in recent years, a clustering technology is used for preprocessing photovoltaic power data, and a targeted prediction model is constructed by classifying different weather conditions, running modes or output forms, so that a technical path comprising K-means clustering, fuzzy C mean clustering, dynamic self-organizing mapping, neighbor propagation, gaussian mixture model and the like is formed. The Deep Convolution Embedded Clustering (DCEC) is focused on the fact that the DCEC can extract spatial structural features from a power day-time matrix by utilizing a convolution network, but the traditional DCEC has obvious limitation that the DCEC mainly captures local spatial correlation of data, only relies on convolution operation to extract local neighborhood textures on the day-time matrix, ignores strong time sequence features of a photovoltaic power sequence, is difficult to capture trend and short-time dependence of time dimension, and can divide a sunny short-time shielding and a typical cloudy day into the same cluster, so that cluster separability is insufficient, further accuracy of a downstream prediction model is affected, and the DCEC is one of core reasons for unstable performance of a traditional cluster-prediction framework in a distributed photovoltaic scene. In the aspect of prediction models, the prior art can be divided into a physical model, a statistical model, a neural network model, a hybrid model and the like, wherein the hybrid model is more suitable for a distributed photovoltaic power prediction scene due to the fusion of various model advantages, and related researches have attempted to improve the prediction performance in modes of matching and combining three-dimensional CNN with similar Yun Kuang, introducing a self-attention mechanism in a two-way long-short-term memory network, combining a time convolution network with BiGRU, constructing a parallel architecture and the like. However, the existing distributed photovoltaic prediction method still faces a plurality of unresolved challenges in terms of model construction, feature mining and meteorological data utilization, namely, the traditional DCEC only extracts spatial features from a one-dimensional structure, ignores strong space-time characteristics of a photovoltaic curve, and causes the problem that 'sunny short-time