COMPARISON OF ACCURACY BETWEEN LSTM AND FEEDFORWARD NEURAL NETWORKS IN PREDICTION OF DAM INFLOW

Abstract

 We conducted a comparative accuracy analysis of LSTM and FFNN models, which are major models in deep learning, for prediction of dam inflows.The dam inflow prediction model aims at long-term forecasting, using only precipitation data as input.For FFNN, we prepared two deep learning models, FFNN(A) and FFNN(B), with different optimization functions. The input data, which consists of precipitation data up to 720 hours prior, was divided into several intervals and time-compressed to enhance accuracy.Among the four dams examined, three were constructed using only basin-averaged rainfall data, while one included point rainfall data to verify its impact.Among the three dams constructed using only basin-averaged rainfall data, FFNN(B) exhibited the highest accuracy. In the case where point rainfall data was added, LSTM showed the highest accuracy.However, it is likely that the model's superiority is influenced not by the presence of point rainfall data, but by its ability to learn from floods of a scale similar to the validation flood.