Ultra-short term prediction of wind power based on multi-resolution closed-loop method

In order to improve the accuracy and stability of wind power prediction, a multi-resolution closed-loop wind power ultra-short term prediction method is proposed. Using historical data with different resolutions, the differential data is generated by the pre-predictor with low, medium and high resolutions, and then the differential data is combined with the post-prediction data into the post-predictor with three resolutions for training. Since the differential data is obtained from the prediction data with different resolutions, the post-prediction data containing the differential data contains inconsistent information between different resolutions, so that the post-predictor receives the inconsistency between different resolutions and minimizes the inconsistency after several closed-loop iterations, and finally outputs wind power prediction with different resolutions. At the end of the paper, simulation experiments are carried out using the data of the 2022 KDD Cup competition, and comparison is made with some mainstream wind power prediction models. It is found that the proposed model has high accuracy and stability.


Introduction
In recent years, due to the depletion of fossil energy such as coal, countries around the world pay more and more attention to the development of new energy, and wind energy, as a leader in new energy, has been concerned by countries around the world.With the increase of installed wind power capacity, the randomness, volatility and intermittedness of wind energy resources have brought great challenges to the safety, stability and economy of power system operation.Accurate wind power prediction is an important means to deal with this problem [1] .
At present, the methods of wind power prediction can be divided into three categories: physical methods, statistical methods and artificial intelligence methods.The physical method is a modeling method using the terrain, roughness, wind turbine parameters and other information of the wind farm.This method does not require a large amount of historical data, but the modeling is complicated and uncertain, and is limited by the accuracy of the model.It is mostly used when the site selection of the wind farm and the historical data are lacking at the beginning of the establishment of the wind farm.Literature [2] establishes a roughness change model to solve the influence of wind speed roughness change, and a terrain change model to solve the influence due to terrain change, and combines the two models to achieve the prediction of wind power.Statistical methods are based on historical time series to predict future series [3] .Common statistical method models include autoregressive model [4] , gray prediction model [5] , Markov chain prediction model [6] , etc.Because the statistical method simply considers the statistical relationship between the series without considering the influence of other factors, the prediction accuracy is not high, and the current use is less.The AI approach is to use artificial neural networks to fit data to form models that can predict future data.Literature [7] uses discrete wavelet transform and sequential convolutional neural network to predict wind power in conjunction with attention mechanism and double-depth Q network.Reference [8] uses combined modal decomposition and a bidirectional long short-term memory neural network optimized by Bayes for prediction.Literature [9] improved the genetic algorithm, and used the improved genetic algorithm to optimize the topology structure and weight parameters of Elman neural network to improve the prediction effect of the network.Based on the traditional LSTM network, literature [10] adds a new gating mechanism to improve the context information modeling capability of LSTM network and improve the accuracy of wind power prediction.Literature [11] proposed an input LSTM network for prediction based on overlay noise reduction autoencoder and batch normalization of depth features.Meanwhile, density-based spatial clustering method with noise was applied to process numerical weather prediction data, and representative samples were selected for training, which improved the efficiency of the model.
The above models have improved the prediction accuracy, but these models are all open-loop prediction, and the stability is not high.Therefore, a multi-resolution closed-loop prediction model is proposed in this paper.The prediction model adopts the multi-resolution method and the structure of the front and back predictor, and the back predictor uses the information of the differential data to predict through the closed-loop feedback.Through differential data and closed-loop structure, the model can predict wind power more accurately in the prediction stage.

Multi-resolution prediction model
Wind power is affected by a variety of factors, so even well-trained open-loop models will have very different predictions in different trainings.The open-loop model is independent in different training, and each training will have different results, which will cause the instability of the model prediction.At the same time, the multi-resolution data used by the proposed model comes from the same data and has high correlation and mutual information.Under the mutual influence of such multi-resolution data, the model will have stable prediction output [15] .The implementation principle is as follows:

Training phase
The left side of Figure 1 shows the flow chart of preparing forecast data.First, select the features you want to input from the raw data.This paper uses the value of SHapley Additive exPlanations (SHAP), a method used to interpret the output of a machine learning model, for feature selection.It assigns an importance value to each feature, indicating how much each feature contributes to the final predicted outcome.According to the degree of contribution to the prediction, historical power, wind speed and pitch Angle are selected as the input features of the model.

Postprediction
( ) ) ) Low, medium and high resolution input and output data( , 1 ,  , 1 ), ( , 2 ,  , 2 ), ( , 6 ,  , 6 ) are respectively formed in the original data set with selected features, where, X for input data, Y for output data; each input and output has two subscripts,t for time series, τ 1 for low resolution (data point interval 60min), τ 2 for low resolution (data point interval 30min), τ 6 for low resolution (data point interval 10min).In order to avoid the shading effect caused by the large numerical difference between the feature data of different dimensions, it is necessary to carry out normalization processing.The power and wind speed are normalized to the maximum and minimum values, and the pitch Angle is sinusoidal.The expression is shown in (1)(2): =   −    −  (1) Since the post-predictor needs the differential data generated by the previous predictor to be supported, the pre-predictor needs to be trained and the differential data generated before training the post-predictor, and then the post-predictor needs to be trained.As shown in the timeline in Figure 1, the pre-predicted data of the training stage is first used to train the three pre-predicted predictors, and then the prediction is made respectively and the differential data is formed.The differential data formation expression is shown in (3): where,FS t,τ for differential signal, the lower corner number represents the number of data points contained in different resolutions in a time step.
After the differential data is generated, the differential data is combined with the post-prediction data, and the differential data is used as a feature input.Since the post-prediction data set combines the differential data generated by the pre-predictor, the post-prediction data set contains inconsistent information between different resolutions, which enables the post-predictor to learn the inconsistency between different resolutions during the training process.These inconsistencies allow the post-predictor to better learn the mapping relationship between wind power data input and output, and can make the output more stable.

Prediction stage
A trained predictor can predict wind power over a period of time using a closed-loop mode.The flow chart is shown on the right in Figure 1.First, the former predictor is used to predict and generate the difference data required by the latter predictor, and the construction formula is shown in (4): where, FP t,τ for differential data of different resolutions. ̂,  ,  ̂,  ,  ̂,  for the data were predicted with different resolutions.Then, the differential data is combined with the post-prediction data of different resolutions to form the post-predictor input data into the post-predictor, and the future wind power is predicted by using the inconsistent information in the data.
After that, the output data of the post-prediction model is combined with the input data of the post-predictor to form a new iterative prediction, forming a closed-loop structure until the end of the iteration.The new differential data will be more consistent than the original differential data, allowing for more accurate predictions.

Example analysis
The experimental data of the multi-resolution wind power short-term prediction model proposed in this paper were selected from the 2022 KDD Cup competition data provided by Longyuan Power Group Co., LTD.The data included 8 feature sequences, each lasting 184 days with a sampling accuracy of 10 minutes, and a total of 26496 sampling data points.The SHAP value was used to select three characteristics with the greatest contribution, namely, historical wind power, wind speed and pitch Angle.The proposed model runs on a laptop with a core i5-11400H CPU, NVIDIA RTX 3050 Ti GPU, and 16GB of memory.

Model evaluation index
In this paper, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are selected as the error evaluation criteria for prediction results.The specific expression is as follows: where, y i for the true value of the i th sample; y ̂i for the prediction value of the i th sample; m for the number of samples.

Predictor selection
The proposed model chooses Long Short Term Memory (LSTM) as the main body of the predictor.The model adopts the structure of 2 layers of LSTM layer, 1 layer of full connection layer and 1 layer of output layer.Because the parameters to be fitted will increase exponentially with more network layers, the proposed model only adopts the 3 layer network structure.

Comparison of different models
For the analysis of the prediction accuracy of the model, this paper uses several groups of models to compare with the proposed model, namely LSTM [14], CNN [13] and RNN [12].These models are the mainstream prediction models at present and are widely used.As can be seen from Table 1, the prediction results of the proposed model may be slightly worse than those of other models at certain resolutions, but on the whole, the proposed model still achieves higher accuracy, which proves that the proposed multi-resolution closed-loop wind power prediction method has indeed improved the prediction accuracy.RMSE decreased by 25.78%, 36.67% and 40.73%, respectively.MAE decreased by 31.44%,45.06% and 50.66%, respectively.After many experiments, the proposed model keeps high prediction accuracy, which shows that the model has high stability in different training.Figure 2 shows the prediction curves of different models at different resolutions.

Conclusion
In this paper, a multi-resolution closed-loop wind power prediction method for ultra-short term is proposed.Through comparative experiments, the following conclusions are drawn: •The proposed model has higher overall prediction accuracy under different resolutions.
•The combined prediction of multiple models and the differential closed-loop structure make the model have high stability.

Figure 2 .
Figure 2. Prediction curves of different models under different resolution data.

Table 1 .
Low, medium and high resolution errors of different models.