CN-121983959-A - Photovoltaic power generation power prediction method based on stacked state frequency memory network and transfer learning

CN121983959ACN 121983959 ACN121983959 ACN 121983959ACN-121983959-A

Abstract

Compared with the prior art, the photovoltaic power generation power prediction method solves the defects of insufficient utilization of frequency domain characteristics, insufficient robustness under data uncertainty and reduced prediction precision when new site data is insufficient. The method comprises the following steps of data acquisition and preprocessing, construction of a three-layer stacked state frequency memory network, training and parameter tuning of a source site model, migration and multi-strategy fine tuning of a target site, and prediction of photovoltaic power generation. The invention realizes controllable reservation and forgetting of multi-frequency components through time domain-frequency domain coupling and gating memory updating at a model structure level, enhances multi-scale characterization capability through three-layer stacking and residual error fusion at a network depth level, reduces the influence of new site data deficiency on prediction precision through source site pre-training and target site multi-strategy fine tuning at an engineering application level, and improves short-term photovoltaic power prediction precision and system robustness.

Inventors

YE XIAO
YANG JINGJING
SHUAI WENQING
CHEN JIANWEI
LIU ZHIBO
WU XINYUAN
An Kangning
LV HAITAO
ZHANG JIAJIA
CHEN XIAOFENG
ZHANG LIJUN
MU FENG

Assignees

中国能源建设集团安徽省电力设计院有限公司

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (6)

1. A photovoltaic power generation power prediction method based on a stacked state frequency memory network and transfer learning is characterized by comprising the following steps: 11 Acquiring power generation historical power data and meteorological data of a source station and a target station photovoltaic power station, performing missing value processing, outlier removal and normalization processing on the data, and constructing an input characteristic sequence and a prediction tag, wherein the source station is a photovoltaic power station for pre-training, and the target station is a photovoltaic power station for power prediction to be migrated and adapted; 12 Constructing a three-layer stacked state frequency memory network; 13 Training a source station model and optimizing parameters; 14 Target site migration and multi-policy fine tuning; 15 And (3) photovoltaic power generation power prediction, namely inputting the processed target site input characteristic sequence into a target site optimal model to perform short-term photovoltaic power generation power prediction, and outputting a predicted power result.
2. The photovoltaic power generation power prediction method based on stacked state frequency memory network and transfer learning according to claim 1, wherein the building of the three-layer stacked state frequency memory network comprises the following steps: 21 Setting a three-layer stacked state frequency memory network comprising a tensor space dimension setting module, a Fourier frequency characteristic construction module, a matrix operation logic design module, a gating structure and a cross-layer residual error fusion structure; 22 A tensor space dimension setting module: Determining tensor space dimension according to input data features to capture time sequence and frequency domain features, and setting an input feature matrix , wherein, For inputting a feature matrix space, T represents a time step, F represents a feature quantity, and the input features comprise historical generating power, sunlight intensity and temperature; 23 A fourier frequency characteristic construction module is set: constructing a frequency characteristic matrix based on Fourier transformation, and inputting the frequency characteristic matrix Frequency domain mapping is carried out to obtain a frequency characteristic matrix Satisfies the following conditions , Wherein, the For a discrete fourier transform operator or an equivalent frequency domain mapping operator, K is the frequency characteristic dimension, Is a frequency domain feature matrix; and performing outer product construction on the input characteristic matrix and the frequency characteristic matrix to form a joint tensor , , Wherein the method comprises the steps of For the input feature vector at time t, As the frequency characteristic vector at the time t, The operation of the outer product is represented, For the F-dimensional real vector space, To represent a K-dimensional real vector space; 24 Setting a matrix operation logic design module and a gating structure, and performing three-layer stacked gating memory updating: For the k-th layer state frequency memory unit, k=1, 2,3, calculate the state forgetting gate at time step t Frequency forgetting door The formula is as follows: , , Wherein, the In order to activate the function, As a matrix of weights, the weight matrix, As a result of the bias term, A hidden state is a time on the k layer; Combined forgetting gate obtained by the outer product of state forgetting gate and frequency forgetting gate , , And calculates the input gate , , Wherein, the As a matrix of weights, the weight matrix, Is a bias term; state frequency memory matrix The updated formula of (c) is given by, , Wherein, the Representing the element-by-element product, As a matrix of weights, the weight matrix, As a result of the bias term, For the state frequency memory matrix of the kth layer at time step t-1, To represent the input gate vector of the k-th layer at time step t, Input feature vectors representing time step t; 25 Setting a cross-layer residual error fusion structure, and carrying out cross-layer residual error fusion and output mapping: Layer 1 output hidden state satisfies , Layer 2 and layer 3 output hidden states satisfy , Wherein, the Representing the output hidden state of the layer 1 at the time step t; Indicating the output hidden state of the kth layer at time step t, To represent the output hidden state of the k-1 layer at time step t; Representing the state frequency memory matrix of layer 1 at time step t; representing a state frequency memory matrix of the kth layer at time step t; in order to activate the function, The final output layer maps the layer 3 hidden state to the predicted power, , Wherein, the The predicted value of the photovoltaic power generation power at the moment t; in order to be in the layer 3 hidden state, In order to output the layer weight matrix, Is an output layer bias term.
3. The method for predicting photovoltaic power generation power based on stacked state frequency memory network and transfer learning as claimed in claim 1, wherein the source station model training and parameter tuning comprises constructing a prediction tag set y from source station photovoltaic power generation power data, splicing weather and environmental parameters into a feature matrix X, and constructing an input feature sequence through a sliding time window Corresponding label Training a three-layer stacked state frequency memory network, optimizing model super parameters by a random search or grid search method, and obtaining source site pre-training parameters by adopting MAE, RMSE, MAPE and R2 as evaluation indexes, wherein the method comprises the following steps: 31 Data set cutting and sample construction, namely constructing a tag set y by taking photovoltaic power generation power data of a source station as a target variable, splicing meteorological and environmental parameters into a feature matrix X, and constructing an input sample according to a sliding time window And corresponding label Dividing the sample into a training set, a verification set and a test set; 32 Initializing model parameters, namely initializing three-layer stacked state frequency memory network parameters according to input dimensions and hidden unit numbers, wherein the parameters comprise weight matrixes and bias items of all gate control units and initial states of the state frequency memory matrixes Wherein Initializing to a zero matrix; 33 Training source station and optimizing super parameters, namely inputting training set into three layers of stacked state frequency memory network, iterating and optimizing parameters through back propagation, and taking mean square error as a loss function , , Wherein N is the number of samples, As a result of the fact that the value, Is the predicted value, and the learning rate, the batch size, the hidden unit number and the residual error fusion coefficient are obtained through random search or grid search Optimizing super parameters; 34 After training, MAE, RMSE, MAPE and R2 are calculated on the verification set or the test set as evaluation indexes, and the source site pre-training parameter set and the normalization parameter with optimal performance are saved for migration initialization of the target site.
4. The photovoltaic power generation power prediction method based on stacked state frequency memory network and transfer learning of claim 1, wherein the target site transfer and multi-strategy fine tuning comprises transferring source site pre-training parameters to a target site model, selecting different parameter freezing and fine tuning strategies according to the target site available historical data scale, and carrying out fine tuning training on target site data to obtain a target site optimal model, comprising the following steps: 41 Assigning the source station pre-training parameter set to a three-layer stacked state frequency memory network of the target station according to the network level one by one, namely copying network parameters of the 1 st layer, the 2 nd layer and the 3 rd layer of the source station to the 1 st layer, the 2 nd layer and the 3 rd layer corresponding to the target station respectively, and obtaining initialization parameters of the target station; 42 Target site data size assessment, statistics of available effective historical training days of target site And set a first threshold value And a second threshold value And meet the following ; 43 Policy a gross trim when When the k-th layer state frequency memory network k=1, 2,3 and all parameters of the output layer are allowed to participate in gradient updating; 44 Policy B bottom layer freezing and high layer fine tuning when When the parameters of the k layer and k=1 and 2 are frozen, only the parameters of the 3 rd layer and the output layer are updated; 45 Policy C linear detection when When the parameters of the k layer are frozen, k=1, 2 and 3, only the output layer weight matrix is updated And bias term ; 46 And (3) terminating the fine tuning training and selecting an optimal model, wherein the fine tuning training adopts a mean square error as a loss function, an early stopping mechanism is set through verification set error monitoring, and when the verification set error reaches a minimum value, the current parameter is saved as the optimal model of the target site.
5. A computer readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the photovoltaic power generation power prediction method based on the stacked state frequency memory network and the transfer learning according to any one of claims 1 to 4 can be implemented.
6. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, which when executed by the processor, implements the stacked state frequency memory network and transfer learning based photovoltaic power generation power prediction method of any of claims 1-4.

Description

Photovoltaic power generation power prediction method based on stacked state frequency memory network and transfer learning Technical Field The invention relates to the technical field of new energy power generation prediction, in particular to a photovoltaic power generation power prediction method based on a stacked state frequency memory network and transfer learning. Background As the scale of photovoltaic installation continues to rise, grid scheduling and power station energy management place higher demands on the accuracy and stability of short-term (minute-scale, hour-scale) power predictions. The photovoltaic output power is affected by factors such as rapid cloud cover change, aerosol and temperature and humidity, wind speed, component temperature effect and the like, and the photovoltaic output power presents strong nonlinearity, multi-time scale and multi-frequency component coupling characteristics, not only comprises low-frequency trends such as daily periods and the like, but also comprises high-frequency random fluctuation caused by cloud cover, and meanwhile, engineering data also commonly has uncertainty problems such as lack of measurement, noise, communication delay, sensor abnormality and the like, so that the prediction difficulty is further increased. The existing method mainly comprises a physical mechanism model, a statistical model and a machine learning/deep learning model. While the depth models such as RNN/LSTM/GRU/transducer have time sequence modeling capability, most of the depth models still use a time domain hidden state as a core, lack of explicit representation and controllable updating mechanism for frequency domain components, cause aliasing of periodic items and sudden high-frequency disturbance of different scales in the same hidden space, and are difficult to simultaneously give consideration to accurate depiction of periodic fluctuation and sudden disturbance, and prediction results are easy to lag or oversmoor and have insufficient robustness under fast-changing weather. In addition, in a newly built power station or station switching scene, the historical data of the target station is less and the missing proportion is high, the target station is easy to be fitted or difficult to converge from zero training, and the existing model is easy to generate domain offset when being directly migrated, so that the performance is obviously degraded under the condition of a small sample. Disclosure of Invention The invention aims to solve the defects of insufficient utilization of frequency domain characteristics, insufficient robustness under data uncertainty (deficiency and noise) and reduced prediction precision when new site data are insufficient in the photovoltaic power generation power prediction method in the prior art, and provides a photovoltaic power generation power prediction method based on a stacked state frequency memory network and transfer learning to solve the problems. In order to achieve the above object, the technical scheme of the present invention is as follows: A photovoltaic power generation power prediction method based on a stacked state frequency memory network and transfer learning comprises the following steps: Acquiring power generation historical power data and meteorological data of a source station and a target station photovoltaic power station, carrying out missing value processing, outlier removal and normalization processing on the data, and constructing an input characteristic sequence and a prediction tag, wherein the source station is a photovoltaic power station for pre-training, and the target station is a photovoltaic power station for power prediction to be migrated and adapted; Constructing a three-layer stacked state frequency memory network; Training a source site model and optimizing parameters; target site migration and multi-strategy fine tuning; And (3) photovoltaic power generation power prediction, namely inputting the processed target site input characteristic sequence into a target site optimal model to perform short-term photovoltaic power generation power prediction, and outputting a predicted power result. The construction of the three-layer stacked state frequency memory network comprises the following steps: The method comprises the steps of setting a three-layer stacked state frequency memory network, wherein the three-layer stacked state frequency memory network comprises a tensor space dimension setting module, a Fourier frequency characteristic construction module, a matrix operation logic design module, a gating structure and a cross-layer residual error fusion structure; the tensor space dimension setting module is set: Determining tensor space dimension according to input data features to capture time sequence and frequency domain features, and setting an input feature matrix , wherein,For inputting a feature matrix space, T represents a time step, F represents a feature quantity, and the input features com