A deep learning model coupled with metaheuristic optimization for urban rainfall prediction

Abstract

• Proposed a deep learning model coupled with optimization for urban rainfall prediction. • Proposed a new optimization algorithm, IWOA, for better turning the model hyperparameters. • Quantitative uncertainty analysis demonstrates high accuracy and low uncertainty of the model. • The model is compared with other models for rainfall prediction. Accurate urban rainfall prediction is crucial for water resource management, flood defense and early warning, and disaster reduction. Leveraging deep learning allows us to establish effective data-driven forecasting models that can facilitate timely flood warnings, optimize water resource allocation, and minimize the impacts of natural disasters such as waterlogging. To address the challenges of urban rainfall prediction, this paper develops an innovative ensemble deep learning model that integrates convolutional neural network (CNN), long short-term memory network (LSTM), and a self-attention mechanism, enhanced by an improved whale optimization algorithm (IWOA). This approach utilizes historical rainfall sequences as inputs to predict future rainfall, thereby simplifying the prediction model’s complexity. The CNN effectively extracts invariant and hidden features from the rainfall data, while the LSTM captures long-term dependencies within the rainfall sequences. Meanwhile, the self-attention mechanism is employed to further capture internal correlations within the rainfall sequence. To rapidly acquire the optimum hyperparameters for the prediction model, the IWOA algorithm incorporates three strategies, including the double random opposition-based learning (DROBL) strategy, the crosswise intersection strategy, and the local search strategy. The proposed DL-IWOA model has been validated for daily and monthly rainfall in Guangzhou City, China, and benchmarked against five other forecasting models. The results demonstrate the model’s superior performance, achieving a NSE value of 0.83 for daily rainfall prediction, with a d-factor of 0.414 and a p-factor of 0.950. For monthly rainfall prediction, the model attains a NSE value of 0.85, with a d-factor of 0.827 and a p-factor of 0.883. By conducting rainfall predictions across multiple sites and comparing the DL-IWOA model’s results with those of five other models, it is found that the DL-IWOA model not only delivers higher prediction accuracy but also exhibits lower uncertainty, underscoring its potential as a valuable tool for urban rainfall forecasting.

Topics & Concepts