CN-115600498-B - Wind speed forecast correction method based on artificial neural network

CN115600498BCN 115600498 BCN115600498 BCN 115600498BCN-115600498-B

Abstract

The invention relates to a wind speed forecasting correction method based on an artificial neural network, which belongs to the field of artificial intelligence and comprises the following steps of S1, sorting 10-meter wind field historical forecasting grid point data and corresponding live data in a target area, preprocessing the data, S2, carrying out clustering analysis on the data by using a clustering algorithm to generate category characteristics, S3, carrying out data characteristic analysis on the 10-meter wind field data by using a time sequence forecasting method to generate a characteristic vector, then adding the category characteristics generated in the step S2 and forecasting timeliness of a sample into the characteristic vector, processing the historical data by using a sliding window method to generate a sequence sample, S4, designing and training an LSTM-based wind speed forecasting correction model, and S5, correcting the wind speed forecasting by using the trained LSTM-based wind speed forecasting correction model.

Inventors

ZHAO XUELIANG
YUAN XIN
ZHANG YULIN
SUN QILONG

Assignees

中国科学院重庆绿色智能技术研究院

Dates

Publication Date: 20260512
Application Date: 20221025

Claims (5)

1. A wind speed forecast correction method based on an artificial neural network is characterized by comprising the following steps: S1, sorting 10 meters of wind field history forecast grid point data and corresponding live data in a target area, and preprocessing the data; S2, carrying out clustering analysis on the data by using a clustering algorithm to generate category characteristics; S3, carrying out data characteristic analysis on 10-meter wind field data by adopting a time sequence prediction method to generate a feature vector, adding the category features generated in the step S2 and the forecast aging of the samples into the feature vector, and processing historical data by adopting a sliding window method to generate a sequence sample; S4, designing and training a wind speed forecast correction model based on LSTM; S5, correcting the wind speed forecast by using the trained LSTM-based wind speed forecast correction model; In the step S1, historical forecast lattice point data containing 10 meters of wind field data in a target area are sorted, lattice point live data corresponding to the historical forecast lattice point data are collected, the historical data span is more than one year, the historical data span comprises forecast hours after forecast time is 48 hours, the forecast time is reported by UTC 00:00 and UTC 12:00, the sorted wind field data are subjected to data analysis, and then analyzed and processed data are stored through a MongoDB database; In the step S2, the data preprocessed in the step S1 are subjected to cluster analysis by using a K-means clustering algorithm to generate category characteristics which are used as one of the characteristics used in the next step, wherein the characteristics comprise statistical characteristics and position characteristics, the statistical characteristics are general characteristics representing historical NWP wind speed errors and additional related information, and the position characteristics are groups clustered by using a K-means unsupervised learning algorithm using geographic position data; In step S3, 8 feature vectors including ori_ diff, max, min, mean, median, Q _25, Q_75 and std are generated, each generated sequence has a length of 30, the number of features is 8, the interval is 12 hours, the span is 15 days, namely the shape of each sample is 30 8, The label corresponding to each sample is the next ori_diff of the sequence.
2. The method for correcting wind speed forecast based on artificial neural network according to claim 1, wherein in step S3, the wind speed forecast process is defined as = (-), Wherein Is the predicted wind speed of 10 meters, Is a numerical prediction model, and when collecting actual measured values, the prediction error The method comprises the following steps: = - In the formula, Is the NWP wind speed at time t, Is the wind speed of 10 meters measured at time t; The wind field data is analyzed by adopting a time sequence prediction method, and specifically comprises the steps of learning and modeling a historical error by developing a model based on previous data and applying the model to predict future values within a period of time so as to predict the future error, and correcting the NWP wind speed by adding the prediction error after predicting the error, wherein the formula is as follows: = = + Wherein the method comprises the steps of Is the prediction error of the time t and, Is the corrected NWP wind speed.
3. The method for correcting wind speed forecast based on artificial neural network according to claim 1, wherein in step S4, training a generalized model with a position mode and a forecast period as input features comprises: Modeling time series data with a cyclic structure by using a long-short-term memory artificial neural network LSTM, modeling a latest past to future target variable of a functional relation between input features, wherein a unit state c and three gates are imported into the LSTM, the gates consist of a sigmoid neural network layer and a point-by-point multiplication operation, and the input gate is used for inputting the current input And previous output Weighting: =σ Wherein the method comprises the steps of And The weight matrix and bias vector, respectively, for input gate, σ () is a sigmoid function with an output range of (0, 1), which tends to select all information in the cell state when approaching 1, and tends to discard the input information when approaching 0; then by multiplying the information selection with the input vector: := Wherein the method comprises the steps of And Respectively a weight matrix and a bias vector of an input layer; the selected information is expressed as: := ⊙ Wherein +.is the multiplication by element, Is the input information combined from the current input and the previous output, Is selected information of an input layer to be stored in a cell state; The forget gate decides which part of the information stored in the cell state should be forgotten, as shown in the following equation: = , Wherein the method comprises the steps of And A weight matrix and a bias vector for gate, respectively; The new cell states are: output gate determines which part of the information should be output, i.e , Wherein the method comprises the steps of And The weight matrix and the bias vector for the gate, respectively.
4. The wind speed forecast correction method based on the artificial neural network as described in claim 3, wherein the method is characterized by comprising the following steps of , , ,..., ) Input into the LSTM network, the LSTM network gives a prediction of the next data, i.e From the slave And extracting ori_diff variables, and correcting the forecast data of the next time.
5. The wind speed forecast correction method based on the artificial neural network of claim 1, wherein the wind speed forecast correction model training process based on LSTM is as follows: dividing the data set into a training set and a verification set; On the training set, obtaining optimal super-parameter configuration by adopting a 5-fold cross validation method, wherein the optimal super-parameter configuration comprises the number of LSTM units, an initial learning rate and a batch size; Training model weights by adopting a random gradient descent method, and setting a maximum training algebra; wind speed data is a continuous variable, using the mean square error MSE as a loss function: training the model until the loss function is the lowest, obtaining optimal model parameters, and finishing training.

Description

Wind speed forecast correction method based on artificial neural network Technical Field The invention belongs to the field of artificial intelligence, and relates to a wind speed forecast correction method based on an artificial neural network. Background With the increasing consumption of traditional energy sources, the research and development of clean renewable energy sources is becoming urgent and important. The wind energy is environment-friendly and renewable, and attracts more and more national investment to establish the wind power plant. The wind speed is a key factor influencing the continuous stability of wind power generation, and therefore, higher requirements are put on wind speed forecast. In recent decades, with the rapid development of Numerical Weather Prediction (NWP), the process of wind speed prediction is changed from a traditional predictor to a subjective and objective combined mode, i.e. an "empirical prediction+nwp result" mode by analyzing weather patterns. Although numerical weather forecast (NWP) is an objective quantitative calculation, the calculation is based on grid points, and the representative meaning is an average value of weather elements in a rectangular area, but element forecast cannot be directly made for an observation station where an weather station is located, which requires a predictor to perform artificial experience correction from grid point to station. When encountering complex weather conditions, the forecaster also determines the final weather element forecasting result in a multi-person consultation mode, so that subjectivity is inevitably brought to a great extent, and human errors which are not easy to quantify are introduced. At present, a plurality of methods suitable for wind speed forecast correction exist, and widely applied methods can be divided into two main types, namely a physical model and a statistical model. The numerical weather forecast (NWP) system is a common physical model for weather forecast and wind speed forecast, however, the physical model is based on a huge, complex and experimental meteorological system, and needs a great amount of meteorological information and physical mechanisms, which makes the application and modeling of the physical model complex. Statistical models require historical data such as wind speed and time stamps to predict, and various statistical learning algorithms are also applied to wind speed prediction industries, such as an integrated moving average autoregressive model (ARIMA) model, an Artificial Neural Network (ANN), and the like, and some hybrid methods are also available. However, these statistical model-based wind speed predictions can achieve good accuracy in short-term predictions, but have poor performance in long-term predictions due to error accumulation problems. Although physical models such as NWP can realize large-time scale prediction, the prediction result is rough and has low accuracy. Disclosure of Invention In view of the above, the present invention aims to provide a wind speed forecast correction method based on an artificial neural network. According to the method, a manual neural network model is trained by using a historical numerical weather forecast result and a live observation value, and then the model is applied to the numerical weather forecast result issued in real time every day to obtain a station-level wind speed forecast correction result, so that the existing forecast product is better utilized, the past forecast experience is inherited, and the uncertainty caused by manual subjective intervention is thoroughly eliminated. The model can reduce the prediction error to the maximum extent, and provides a high-precision, pure and objective real-time wind speed prediction post-processing method in wind speed prediction. In order to achieve the above purpose, the present invention provides the following technical solutions: a wind speed forecast correction method based on an artificial neural network comprises the following steps: S1, sorting 10 meters of wind field history forecast grid point data and corresponding live data in a target area, and preprocessing the data; S2, carrying out clustering analysis on the data by using a clustering algorithm to generate category characteristics; S3, carrying out data characteristic analysis on 10-meter wind field data by adopting a time sequence prediction method to generate a feature vector, adding the category features generated in the step S2 and the forecast aging of the samples into the feature vector, and processing historical data by adopting a sliding window method to generate a sequence sample; S4, designing and training a wind speed forecast correction model based on LSTM; s5, correcting the wind speed forecast by using the trained LSTM-based wind speed forecast correction model. In step S1, historical forecast lattice point data including 10 meters of wind field data in a target area is sorted, lattice point live data