Rainfall prediction using ML, CNN+LSTM, and transformer models

Everything depends on the weather, be it a farming calendar, deciding on whether to take an umbrella or to smooth away against the floods, the weather is involved. In the given research project, I examined what the historical ML methods could look like compared to the deep learning in the situations where all we need to determine is whether it is raining or not. My data was directly out of the weather stations and consisted of all of the standard names of temperature, humidity, pressure, wind speed, and cloud cover as well as some convenient engineered features such as seasonal indicators and lagged variables. I divided the data into training and validation as well as a test set and I addressed the imbalance of the classes by using SMOTE and achieved hyperparameter optimization by using GridSearchCV. The models used on the ML side of the fence as my benchmarks included Logistic Regression, Random Forest, Gradient Boosting and XGBoost. For the deep learning person, I ran on Long Short-Term Memory (LSTM) stand alone, CNN+LSTM hybrid and Rainformer is a transformer model designed for time series prediction. The findings revealed that the older methods of machine learning do not give out till the end, they can be trained much faster and tend to achieve relatively equal, or even higher, levels of accuracy in comparison to deep learning cycles. But the CNN+LSTM hybrid was the one that did best overall, with an accuracy of 84 percent and a very high F1-score. 89. Two percentage points higher than the others. This trick demonstrates the benefit of hybrid models in learning spatial and temporal features in data. In conclusion, the study has demonstrated that although the traditional methodology may still be very competitive, an ensemble of DNNs could drive rainfall forecast into the future and enable operational weather systems.