Multi-Scale Weather Forecasting Using Deep Learning Architectures With Chennai Climate Data

Weather forecasting is an essential aspect of climate-sensitive industries like agriculture, water resources management, and disaster risk management. Short-range forecasts enable prompt decision-making, whereas medium-range and long-range predictions are vital for strategic decision-making and policy formulation. Traditional forecasting models tend to fail to detect the intricate, non-linear, and scale-dependent processes inherent in meteorological records. Although classical models are capable of providing some level of predictability, they tend to lack the ability to describe temporal dynamics and nonlinear relationships of meteorological information. Deep learning is increasingly becoming an influential alternative because it has the capability for modeling sequence dependencies and spatiotemporal patterns. This research deals with the issue of enhancing multi-scale weather forecasts using sophisticated neural architectures. Particularly, it compares and examines LSTM, LSTM-CNN, and LSTM-Transformer models for forecasting temperature, humidity, and rainfall at different time resolutions. Weather data for Chennai between 2000 and 2025 were retrieved through the Open-Meteo API. The data were resampled into daily, weekly, and monthly scales, normalized, and fed into a walk-forward validation process. Each model was tuned with Keras Tuner and evaluated using different metrics. Findings indicate that LSTM-CNN has the best performance for short-term forecasting because it can learn local patterns, and LSTM-Transformer is best suited for long-range forecasting with global attention mechanisms. Rainfall, because it is bursty, still proves to be the hardest parameter to accurately model. The research finds that architecture choice should be dependent on the forecast horizon, and that hybrid models are promising candidates for improving accuracy and scalability. These results support the creation of intelligent, geographically specific climate forecasting systems for climate-resilient decision-making. These insights can directly support agricultural scheduling and water resource planning, offering region-specific decision support for climate resilience.