Daily reservoir inflow forecasting using multiscale deep feature learning with hybrid models

• A multiscale deep feature learning method with hybrid models is proposed to forecast daily inflow.
• Both EEMD and FT techniques are applied for multiscale feature extraction.
• A DBN is used as a deep feature learning approach.
• Hybrid D-NNs are employed for features forecasting.
• This method improves inflow forecasting accuracy due to the capacity of understanding sophisticated features sufficiently.

SummaryInflow forecasting applies data supports for the operations and managements of reservoirs. A multiscale deep feature learning (MDFL) method with hybrid models is proposed in this paper to deal with the daily reservoir inflow forecasting. Ensemble empirical mode decomposition and Fourier spectrum are first employed to extract multiscale (trend, period and random) features, which are then represented by three deep belief networks (DBNs), respectively. The weights of each DBN are subsequently applied to initialize a neural network (D-NN). The outputs of the three-scale D-NNs are finally reconstructed using a sum-up strategy toward the forecasting results. A historical daily inflow series (from 1/1/2000 to 31/12/2012) of the Three Gorges reservoir, China, is investigated by the proposed MDFL with hybrid models. For comparison, four peer models are adopted for the same task. The results show that, the present model overwhelms all the peer models in terms of mean absolute percentage error (MAPE = 11.2896%), normalized root-mean-square error (NRMSE = 0.2292), determination coefficient criteria (R2 = 0.8905), and peak percent threshold statistics (PPTS(5) = 10.0229%). The addressed method integrates the deep framework with multiscale and hybrid observations, and therefore being good at exploring sophisticated natures in the reservoir inflow forecasting.