Sample Entropy based Variational Mode Decomposition with Hybrid RNN for Short Term Wind Power Interval Prediction

Abstract

Intervals in wind energy predictions are an excellent way to quantify uncertainty. Wind power's highly variable nature makes it challenging to achieve good-quality prediction intervals (PIs). The Lower Upper Bound Estimation (LUBE) method is commonly used in interval prediction. However, the existing LUBE technique is trained either using shallow statistical models or rudimentary profound learning models that restrict its capability. As a result, the authors of this paper choose to combine the LUBE method with two hybrid models, namely CNN-LSTM (Convolutional Neural Network-Long Short Term Memory) and BiLSTM (Bidirectional LSTM). A developed interval-based optimization strategy with an improved cost function was used to highlight the advantages of these two networks. This improved cost function takes into account the location disparity between prediction intervals and constructed intervals, resulting in better control over PICP (Prediction Interval Coverage Probability) and PINRW (Prediction Interval Normalized Root Mean Squared Width), ensuring better adjustment capability. The suggested CNN-LSTM and BiLSTM algorithms were compared to the performance of other deep learning models on two different datasets that differed geographically. To reduce the data's complexity, it was treated with a noise-free procedure known as VMD (Variational Mode Decomposition). To break down the data and pick subseries, Sample entropy was used. The CNN-LSTM model beat other models in multiple experiments and provided a narrower prediction band with a high coverage probability. According to the results, hybrid models also had a longer run time and took longer to train than non-hybrid models.