Ultra-short term wind power prediction applying a novel model named SATCN-LSTM

Highlights

• The proposed self-attention temporal convolutional network is applied to enhance feature extraction of wind power.

• The meteorological factors are considered in ultra short-term wind power prediction.

• The proposed method can improve the accuracy of wind power prediction by experiments.

Abstract

Accurate and reliable wind power forecasting has become very important to power system scheduling and safely stable operating. In this paper, a novel self-attention temporal convolutional network (SATCN) is combined with long-short term memory (LSTM) to forecast wind power for guaranteeing the continuous electricity supply. In the proposed SATCN-LSTM model, the structure of SATCN with a self-attention mechanism is conducted to pay more attention to features that contribute more to the output. The strength of SATCN is performed through extracting temporal feature of meteorological data and correlation characteristics between variables. LSTM is used after SATCN to further build the connection between features and outputs for predicting future ultra-short time wind power. The effectiveness and advancement of the proposed method is tested by using meteorological data and wind power data from two different wind farms in the U.S. The experimental results reveal that the SATCN-LSTM model is more accurate comparing to other methods. Taking California's fourth quarter wind power forecast results as an example, the proposed method has carried out a reduction of 17.56%, 10.99%,11.34% and 3.68% on the root mean square error compared with LSTM, TCN, CNN-LSTM, TCN-LSTM.

Introduction

For coping with the problems of energy depletion and environment pollution, the energy structure needs to be changed urgently and renewable energy has become a focus of attention. Wind power with its environmentally friendly and renewable characteristics, has been one of the most rapid developing energy sources all over the world[1]. Nevertheless, the volatility and intermittency of wind power generation limit the large-scale integration of wind turbines into the power system [2], which will affect the power quality and steady power system operating. Therefore, the accurate prediction for wind power has important practical value and significance for the power systems.

Many researchers have been committed to exploring accurate wind power prediction models for a long time. Previously, traditional machine learning methods had many applications for sequence prediction because of its small data requirements, simple structure and low hardware requirements. Li et al.[3] utilized improved support vector machine (ISVM) to forecast short-term wind power. In [4], a method applying extreme learning machine (ELM) was provided for predicting wind power. Zhang et al. [5] established a hybrid model which autoregressive moving average (ARMA) was used to predict linear part and principal component analysis-radial basis kernel function (PRBF) was used for the nonlinear part after applying variational mode decomposition (VMD). These used models are hard to express the complicated nonlinear relationship between input and output which limits their prediction accuracies [6]. With the development of big data, the advancement of hardware facilities and technology, deep learning algorithms have been developing rapidly, which provides a novel way in time sequences prediction. On the one hand, LSTM has many applications in wind power prediction as a special type of recurrent neural network (RNN) which is one of the most commonly used models for dealing with time series. In [7] by Zhu et al, LSTM was utilized to forecast wind power and the prediction result of LSTM is more accurate and error is smaller comparing with traditional machine learning methods. M. Shahid et al.[8] utilized genetic algorithm (GA) for optimizing window size and number of neurons of LSTM, which improved predicting accuracy to a certain extent comparing with ordinary LSTM. On the other hand, convolutional neural network (CNN) is utilized in time series prediction. Harbola et al. [9] built one dimension CNN architectures to predict wind speed and emphasized its feature extraction ability. Yildiz et al. [10] converted features which obtained after VMD process into images and employed an improved residual-based deep CNN to predict wind power. The designed CNN produced superior results across compared current CNN deep learning architectures. However, CNN only considers ordinary feature extraction and ignores the characteristics of the time series modeling. Bai et al. [11] redesigned sequential CNN and proposed temporal convolutional network (TCN) which considered characteristics of time series is more suitable for solving time series forecasting problem. Huang et al. [12] performed TCN processing data in parallel, which provided a flexible receptive field. This receptive filed has a robust gradient as the predictor to forecast wind power or wind speed and can achieve better prediction results. The significance of meteorological variables also was demonstrated for wind power prediction [13]. Li et al. [14] applied a hybrid decomposition model and temporal convolutional network for wind speed prediction.

Because of the complexity in time series forecast and limitations of single models, some researchers tried to add feature extraction part before using RNN network to improve forecasting performance. Li et al. [15] designed a new hybrid model in which CNN was used to extract features form input variables and LSTM was utilized to process the features for predicting energy structure of China. The effective performance of the model was verified by comparison with the prediction results of single models. In [16], CNN-LSTM and ConvLSTM were introduced to predict power production for self-consumption PV plants. The study demonstrated that the two models had the ability to model the global properties in a good way and they outperform the LSTM model in predicting power production. Hong et al. [17] utilized CNN to carry out feature extraction of volatile wind power time-series and radial basis function neural network (RBFNN) was cascading after it to predict day-ahead wind power. Chen et al. [18] applied CNN-LSTM to predict wind speed and made a conclusion that the combined model can deeply extract the temporal and spatial correlation features at the same time. For further improving the performance of the model, attention mechanism which can automatically learn and calculate the contribution of inputs to outputs so that the model can focus on important parts is gradually introduced into network structures. Zeng et al [19] developed a DARNN (deep attention residual neural network) model for remaining useful life prediction and the extraction capability and prediction performance of this mode were testified. In [20], a bilinear self-attention CNN model was presented to identify leukocyte type and acquired the highest score in three evaluation indictors and the results proved the proposed model had the best performance.

In this paper, TCN is introduced for extracting the temporal and related features from input data. Furthermore, self-attention mechanism is used to improve the performance of TCN, which can focus on those features that contribute more. Then SATCN is proposed for extraction of the temporal characteristics of the meteorological information and correlation characteristics between variables. LSTM is cascaded with SATCN to further build the connection between features and outputs. The full-year meteorological data and wind power data in 2012 from California and Tennessee are utilized as confirmatory experimental dataset. These data are divided into four quarters for modeling and forecasting. The experimental results present that the proposed SATCN-LSTM method possesses higher accuracy and better applicability than other methods.

The key innovations in this work are manifested below:

• Considering the complex nonlinear relations between meteorological factor and wind power, TCN model is developed for extracting the temporal and spatial features of raw data.

• SATCN is proposed to enhance feature extraction, where self-attention is added so that the model can assign different weights comparing other models.

• A new wind power predicting model called SATCN-LSTM is proposed, which can promote the stability and economic efficiency of power system.

The rest of this research manuscript has been arranged as follows. Section 2 expounds the basic methods about TCN, LSTM and self-attention in brief. The proposed novel model structure and comparison models and the error indicators are described in Section 3. Data description and data preprocess are presented in Section 4. The experimental prediction results for diverse methods are demonstrated in Section 5 and the conclusions are presented in Section 6.