🤖 AI Summary
To address insufficient modeling of seasonality and periodicity in Chennai’s climate time-series forecasting, this paper proposes a novel hybrid method integrating adaptive period estimation, unsupervised learning, and spline interpolation. The method first introduces an unsupervised clustering– and spectral analysis–based algorithm to estimate dominant periods directly from data, eliminating reliance on predefined periodic assumptions. Subsequently, it constructs a spline-enhanced ensemble time-series model that jointly captures trend, periodic, and residual components. Evaluated on a multi-source climate dataset from Chennai, the approach achieves an average 23.6% reduction in MAE over ARIMA, Prophet, and LSTM baselines. It notably improves long-horizon forecast accuracy and cross-seasonal robustness. By combining interpretability with minimal dependence on domain-specific prior knowledge, the framework establishes a new paradigm for regional climate forecasting.
📝 Abstract
This article explores a novel approach to time series forecasting applied to the context of Chennai's climate data. Our methodology comprises two distinct established time series models, leveraging their strengths in handling seasonality and periods. Notably, a new algorithm is developed to compute the period of the time series using unsupervised machine learning and spline interpolation techniques. Through a meticulous ensembling process that combines these two models, we achieve optimized forecasts. This research contributes to advancing forecasting techniques and offers valuable insights into climate data analysis.