Performance comparison of Probabilistic and Artificial Neural Network Models for Long-sequence Generation of Wind Speed Forecasts

Abstract

This paper presents a new method for generating long-sequence wind speed time-series forecasts for purposes of offshore wind farm asset and operations planning. Our goal is to develop a planning decision support tool with which wind farm planners and operators can make informed decisions for development and operation of future offshore wind assets considering revenue and power generation yield, as well as operation and maintenance expenditures.The proposed methodology should be computationally efficient and should be able to reliably generate accurate wind speed time-series forecasts for the required planning timescale. In this paper, we used an Autoregressive Moving average model as benchmark to evaluate and compare the performance of four different artificial neural network models namely, uni-variate Long-Short Term Memory (LSTM), uni-variate hybrid one dimensional convolutional neural network with LSTM (1D-CNN-LSTM), multivariate Long-short term memory and multivariate hybrid 1D-CNN-LSTM architectures. The performance evaluation is delivered through the statistical comparison of the metrics, RMSE, MAE and MAPE, and the final selection of the outperforming model is supported using Diebold-Mariano statistic test. Experiments consists of applying different types of pre-processing to the wind speed dataset and the modification of models’ architecture to include either a batch normalization or a drop-out regularization layer are realized to aid in the selection of the most suitable model engineering. Results suggest the uni-variate hybrid 1D-CNN-LSTM is able to deliver short-term prediction for longer timescales while maintaining a suitable degree of accuracy.