Multivariate CNN-LSTM model for wind power forecast and input variables correlation analysis based on SHAPLEY values

Abstract

Wind Power forecast based on the use of modern Deep Neural Network (DNN) methods is being explored as a tool to support decision-making by wind farm developers and operators for investment, grid integration and operational planning of the site. Techniques such as Long Short Term Memory (LSTM) have shown success in wind prediction for short to midterm planning timescales, however, when predicting wind power generation, input variables can influence the output differently [1]. It has been observed that DNN models do not predict sudden events like failures or outages with the sufficient accuracy due to the randomness attributed to the events [2, 3]. Correlation analysis of the input variables with the one being predicted are normally carried out using well-known statistical methods such as the Pearson correlation estimation, however, time series like wind, wave height or air density are not stationary and may have inherent influence on the wind farm operation in terms of variations in the wind power output. Therefore, in this paper, a weather correlation analysis based on the Shapley value estimation is presented as a way to improve accuracy of wind power prediction models . Furthermore, a multivariate Convolutional Neural Network together with a Long-short Term Memory (CNN-LSTM) model is proposed for short-term wind power forecast suitable for wind farm operational planning applications (e.g., scheduling maintenance routines, and/or wind farm energy/revenue estimation). Hourly to daily seasonal sets of predictions are carried out for a whole year sequence in order to compare and identify the contribution of the wind speed, direction, ambient temperature, and pressure to the power generation prediction taking into account the seasonality of wind. Finally, the accuracy of the proposed prediction model is evaluated by estimating a set of well-established summary statistics namely, the Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) residual analysis whilst comparing to the values obtained from a classical statistical Autoregressive Integrative Moving Average (ARIMA) model. For figures illustrating the architecture proposed and preliminary results, please refer to the supporting document. Figures 1 and 2 show some preliminary results obtained from the model, and figure 3 illustrate the contribution of each feature to the prediction in a particular time stamp.