Transfer learning for reservoir operation based on regional model and large-scale dataset

A regional spatial transfer learning method is proposed to map the attributes of reservoirs and their operating decisions, addressing the issue of poor transfer-ability of reservoir operating models. Firstly, a dataset of reservoir attributes and operating data is constructed, including a total of 503 reservoirs in China and the United States. The 503 reservoirs are divided into source reservoirs (rich data reservoirs) and target reservoirs (data-scarce reservoirs) based on a five-fold cross-validation method. A regional spatial transfer learning method is established to map the attributes of reservoirs and their operating strategies. After training the model network parameters, a generalized reservoir operating model is formed. The information of data-scarce reservoirs is input into the generalized reservoir operating model to achieve adaptive dynamic transfer of reservoir operating strategies. Deep learning interpretability techniques are used to analyze the relationship between static attribute features of reservoirs and the performance of transfer learning. Results show: (1) The generalized reservoir operating model, which maps the static attribute features of reservoirs to reservoir operating decisions, can achieve spatial transfer of operating decisions. After five-fold cross-validation, the average Nash-Sutcliffe Efficiency (NSE) for reconstructing operating decisions of data-scarce reservoirs is 0.69. By evaluating the deviation between the transferred operating decisions and real-world operating decisions of the 503 reservoirs, it is found that 337 reservoirs, accounting for 67.0% of all reservoirs, have an NSE higher than 0.6. (2) Reservoir attribute features can enable the model to explore the relationship between operating decisions of reservoirs with similar features. Among the 503 reservoirs, 288 reservoirs show better transfer learning performance considering static attribute features. (3) Deep learning interpretability techniques are used to analyze the relationship between the transfer learning performance and the static attribute features of reservoirs. The most important factors affecting the transfer learning performance are long-term average discharge at dam location, longitude of point location of dam, area of upstream catchment draining into the reservoir, and maximum storage capacity of reservoir, which explain 31.1%, 11.3%, 10.0%, and 7.2% of the transfer learning model's performance, respectively.