Precipitation forecasting based on deep learning strategy using empirical wavelet transform, Markov chain-incorporated long-short term memory network

Abstract Accurate and stable precipitation forecasting can better reflect the changing trend of climate and also provide timely and efficient environmental information for a management decision, as well as prevent the occurrence of floods or droughts. This study proposes a hybrid model for precipitation forecasting and demonstrates its efficiency. Firstly, the empirical wavelet transform (EWT) is introduced to decompose and pre-analysis hidden characteristics of the precipitation data. Secondly, The Long Short Term Memory (LSTM) network is improved in combination with the Markov Chain (MC) algorithm, thus providing more precise forecasting results for rainless and rainy months and mitigating any extreme and non-physical precipitation generation. Thirdly, the multi-step prediction is explored to improve the reliability and flexibility of rainfall. Monthly precipitation data is used as illustrative cases to verify the performance of the proposed model. Parallel experiments using non-decomposing models, other traditional machine learning approaches optimized by the mind evolution algorithm have been designed and conducted to compare with the proposed model. Results indicated that the proposed hybrid model can capture the nonlinear characteristics of the precipitation time series, thus provides more precise forecasting results.