| Optimised Deep Convolutional Spiking Neural Network for Accurate Long-Term and Short-Term Rainfall Forecasting in Climate Prediction Systems | |
|
MSRDG International Journal of Computer Scientific Technology & Electronics Engineering
© 2026 by MSRDG IJCSTEE Journal
Volume 2 Issue 4
Year of Publication: 2026 |
Paper Download Article ID MSRDG-IJCSTEE-V2I4P101 DOI https://doi.org/10.66037/MSRDG-IJCSTEE/V2I4P101 |
|
Abstract: Reliable rainfall forecasting across both short-term (hourly to daily) and long-term (weekly to seasonal) horizons remains a central challenge for climate prediction systems, hydrological planning, disaster risk reduction, and agricultural decision-making. Conventional deep learning architectures such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and their hybrids have improved forecast skill relative to classical statistical and physics-based numerical models, but they remain computationally expensive and are not naturally suited to representing the discrete, event-driven, temporally sparse nature of precipitation processes. This paper proposes an Optimised Deep Convolutional Spiking Neural Network (DCSNN) that combines the hierarchical spatial feature extraction capability of convolutional layers with the temporally sparse, energy-efficient, event-driven computation of leaky integrate-and-fire (LIF) spiking neurons, and tunes the resulting architecture's hyper-parameters using a metaheuristic optimisation strategy. The proposed framework encodes historical rainfall and auxiliary meteorological variables into spike trains, processes them through stacked convolutional spiking blocks, aggregates temporal spike activity, and produces both short-term and long-term rainfall forecasts from a shared, jointly optimised backbone. We describe the architecture, the spike-encoding and surrogate-gradient training procedure, and the optimisation strategy in detail, and outline an experimental protocol using publicly available rainfall and reanalysis datasets. Representative results illustrate the type of accuracy, error-reduction, and convergence behaviour expected of the framework relative to SVR, Random Forest, LSTM, ConvLSTM, and CNN-BiLSTM baselines. The discussion highlights the trade-offs between spiking-network energy efficiency and training stability, and outlines directions for validating the framework on real, station-level and gridded rainfall datasets. |
|
| Keywords: Spiking neural network, deep convolutional neural network, rainfall forecasting, climate prediction;, metaheuristic optimisation, short-term forecasting; long-term forecasting, time-series prediction | |
