Synthetic Scenario Generator and Forecasting Evaluation Framework for Volcanic Air-Quality Time Series

This repository offers a comprehensive and reproducible approach for assessing time-series forecasting models in both real-world and synthetic non-stationary contexts pertaining to volcanic air quality monitoring.The application was created for a research report submitted to PeerJ Computer Science about adaptive forecasting for volcanic air pollution in Indonesia. It facilitates scenario-aware assessment of forecasting models during various stages of volcanic activity, emphasizing practical AI implementation instead of introducing novel techniques.The repository has the following essential components:• Data Acquisition and Preprocessing: Air quality and weather data from OpenWeather services are collected through an API and organized into combined time series for different cities in Indonesia.• Generation of synthetic scenariosControlled non-stationary scenarios are produced by a specialized scenario-generating module that implements perturbations, including random noise injection, sudden spike insertion, regime transitions, recovery trends, and prolonged prediction horizons.• Integration of forecasting modelsThe experimental pipeline integrates eleven forecasting models, comprising ARIMA, recurrent neural networks (Simple RNN, GRU, Bi-LSTM), convolutional models (1D-CNN, CNN-LSTM, TCN), attention-based architectures (Transformer), and generative models (DeepAR, N-BEATS, LSTM-VAE).• Assessment and representation of performance Model performance is assessed using standard regression metrics, including MAE, RMSE, and R², accompanied by comparison visual outputs enabling scenario-based analysis.