Large-scale hydropower dispatching system based on cloud platform and its key technologies

Abstract

In the past decade, the rapid expansion of hydropower capacity in China brings challenges to hydropower dispatching in data management and operation scheduling. In data management, the scale of hydropower operation data increases explosively, presenting typical big data features. In operation scheduling, the optimization of hydropower system operations is a high-dimension, nonconvex, and nonlinear problem. Conventional optimization methods encounter the “dimensionality disaster”, and cannot solve the scheduling problem efficiently. To deal with these challenges, by analyzing the data characteristics and operation scheduling requirements, this paper develops a large-scale hydropower dispatching system based on the cloud platform. The dispatching system consists of four layers: data access layer, cloud platform layer, data sharing layer, and application layer, which has good scalability and flexibility in data management and operation scheduling. In the application layer, key technologies based on the big data and cloud computing resources of the dispatching system are proposed to carry out the optimal operation of hydropower systems, including distributed parallel dynamic programming (DPDP) for long-/medium-term scheduling and data mining-based method (DMM) for short-term scheduling. Finally, the dispatching system is applied to the operation of hydropower systems in southwest China. Results indicate that: 1) The developed dispatching system is capable of processing 18 million data records per day, with a peak processing speed of 10,000 data requests per second, fulfilling the requirements for hydropower dispatching. 2) In long-term generation scheduling, under varying levels of runoff and plant scales, DPDP has demonstrated a remarkable reduction in computing time compared to dynamic programming, obtaining the same optimal solution with over 90 % less computational time. 3) In short-term peak shaving operation, on typical dry and wet days, the computing time of DMM is 7.2 s and 7.5 s, respectively, far less than the time of mixed-integer linear programming, and the maximum residual load obtained by DMM is lower than historical observations, indicating its effectiveness.