A Hybrid Time-Series Forecasting Model for Long-Term Climate Pattern Prediction Using ARIMA and Deep Learning Techniques

Abstract

This project provides a web-based hybrid time-series forecasting tool designed to improve the accuracy of long-term climate forecasts using “statistical and machine-learning methods. The tool captures historical climate data with geographic coordinates and measurements and temporal features such as year and month. Data cleaning, transformation, encoding, and normalization are conducted to preserve the quality of the data and help build users’ trust in the system. This tool has a variety of forecasting methods available. To learn from features, Random Forest Regression is used. To model linear trends over time, the Autoregressive Integrated Moving Average (ARIMA) method is used. To learn and model relationships which are not linear in the data, Long Short-Term Memory (LSTM) is used to model the residual data. This tool uses the outputs of the previously mentioned systems to improve the accuracy of the forecasts. The application was developed using the Flask framework to build a web tool to allow users to secure an account, upload data, train the model, and make forecasts. Data visualization methods such as histograms, heatmaps, and performance comparison visualization are used to provide analysis and help interpret the results. The predictions generated are evaluated using the R-squared method. This tool provides a simple and powerful way to forecast climate data and supports planning and management of the environment using the data