A CNN encoder decoder LSTM model for sustainable wind power predictive analytics

Abstract

Wind Power (WP) proliferates as one of the significant sustainable energies available in the form of temporal intervals. WP exists as a natural energy generating resource that depends on climatic parameters like wind speed, wind direction. These parameters are highly undetermined, thus collected over a fixed time interval. Hence, the nature of WP is quite uncertain over the time sequences and intricate to forecast. Accurate WP prediction has high significance for wind plants. It sequences the planning for different types of wind farm layouts, number and type of wind turbines required, peak load estimation, distribution areas and plans, cost and benefit risk mitigation etc. Long Short-Term Memory (LSTM) involves prediction using temporal relationships of data points collected over a period of time. The type of predictive analytics required in WP forecasting is very similar to the working of LSTM model. A hybrid Deep Learning (DL) based model using encoder decoder framework with Convolution Neural network (CNN) encoder with LSTM as decoder, named CNN Encoder-Decoder LSTM (CNN-ED-LSTM) is proposed in this study. Hence, to enhance the existing WP predictive analytics (WPPA), this paper focuses on optimization of WPPA using LSTM as the decoder and one-dimensional CNN as encoder. Evaluation of the proposed model is done using five different performance evaluation metrics Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE) and computational time. Also, other traditional DL models such as VanillaLSTM, StackedLSTM, CNN-LSTM and Bidirectional-LSTM (Bi-LSTM) are assessed over the same metrics. Additionally, five statistical models namely, ARIMA-LSTM, AutoARIMA, AutoRegressive Integrated Mean Average (ARIMA), AutoRegressive Mean Average (ARMA) and AutoRegressive (AR), are evaluated. Then, the efficacy of CNN-ED-LSTM is compared with these conventional models of time series predictive analytics as well as traditional DL models that are already used in different studies for WPPA. The proposed CNN-ED-LSTM model for WPPA is performing better than traditional DL models from 3.2% to 9.6% in terms of MSE metric and 3.2–7.5% for RMSE metric.