UNRAVELING PATTERN AND FORECASTING URBAN RAINFALL USING TIME SERIES ANALYSIS

The primary objective of this urban study is to identify the most effective forecasting method for highly seasonal time series data, using monthly rainfall records for Chennai from 1901 to 2021. The analysis begins with data visualization to uncover long-term trends and seasonal variations. We apply clustering techniques specifically to seasonal components of the rainfall data to group similar seasonal behaviours and reveal distinct rainfall regimes across different periods. The structure and distribution of data within each cluster are analyzed to better understand rainfall variability and recurring seasonal patterns. Following this, three forecasting models-ARIMA, STL decomposition, and seasonal naïve forecasting-are implemented. The performances of these methods are evaluated using the standard metrics of Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Scaled Error (MASE). Among the models tested, STL decomposition performs the best, achieving the lowest MAE (67.99), RMSE (125.03), and MASE (0.67). Its ability to isolate trend, seasonality, and residuals allows for more accurate forecasting of complex and highly seasonal rainfall patterns. These findings demonstrate the value of integrating clustering with seasonal analysis and underscore the robustness of STL decomposition in environmental time series forecasting. Leveraging this finding, STL decomposition is utilized to forecast rainfall for the entire dataset. Forecasted values are merged with the original data to reapply K-means clustering and validate consistency in rainfall regimes. The analysis reveals a remarkable similarity in the distribution of data across the new clusters, indicated by an Adjusted Rand Index of 0.95.This shows that STL decomposition has effectively captured the underlying trends and patterns in this highly seasonal data.