Leveraging a deep learning model to improve mid- and long-term operations of hydro-wind-photovoltaic complementary systems

The intrinsic high variability in wind and solar power can cause instability in national energy systems. Thus, wind and solar power are usually integrated with hydropower systems, forming Hydro-wind-photovoltaic complementary systems (HWPCSs), to accommodate such highly variable renewables. However, effective operations of HWPCSs are often limited by the mismatch of operational decisions across different horizons (short-, mid-, and long-term), stemming from the approximations in mid- and long-term power outputs. To address this issue, this study proposes to leverage a deep learning model (DLM) to improve mid- and long-term operations of HWPCSs. The DLM is established by: (1) generating a large set of mid- and long-term operation samples that incorporate details of short-term operations; (2) training and validating the DLM to capture nonlinear relationships between inputs and power outputs; (3) identifying dominant inputs with the well-trained DLM; and (4) applying the DLM to realistic operation. The JP-Ⅰ HWPCS is selected as the case study. Results show that: (1) the DLM has high accuracy and robustness in estimating mid- and long-term power outputs (R2 > 0.992, NRMSE<0.022, MAPE<5.55 %, and ARE<0.71 %); (2) the (initial and terminal) reservoir water level and inflow dominate the total power outputs, exceeding the effects of inputs for wind and photovoltaic power; (3) compared to the conventional method, the DLM helps to increase the total power output by 6.81 % during a typical year. Overall, the DLM captures the connections across short-, mid-, and long-term operations well, driving the optimized decisions of the overall HWPCS closer to the theoretically optimal solution.