Abstract

The capability to predict rainfall effectively is crucial for both economic planning and safety measures. This project examined the performance of three well-known forecasting models: Long Short-Term Memory (LSTM), Autoregressive Integrated Moving Average (ARIMA), and Seasonal Autoregressive Integrated Moving-Average (SARIMA) to determine their accuracy in predicting rainfall patterns. Extensive analysis of data was conducted to identify which model was the most reliable and accurate, considering varying climatic conditions and time scales. The LSTM model, a type of network designed for sequential data, was expected to excel due to its ability to understand long-term dependencies in data series. This is vital for decoding meteorological data influenced by complex physical and time-based dynamics. The architecture of LSTM enabled it to leverage vast amounts of historical rainfall data, allowing it to grasp the subtleties and complexities of weather patterns more effectively than its competitors. Conversely, the ARIMA model, which predicts future points based on past values and error corrections, fails to naturally incorporate seasonality—an essential factor in rainfall forecasting. This gap could result in poorer performance, especially in regions experiencing clear seasonal weather variations. Although robust for various forecasting tasks, ARIMA was predicted to fall short in the rainfall context because of its limitations in modeling seasonal trends. Meanwhile, SARIMA, which builds on the ARIMA model by adding seasonality, was more suited for this task. It was anticipated to achieve moderate success by adjusting for seasonal fluctuations in rainfall data. SARIMA's capacity to handle both seasonal and non-seasonal elements might render it more effective than ARIMA, yet it might still lag behind LSTM, which can manage complex non-linear relationships due to its advanced deep learning capabilities. This investigation employed an exhaustive dataset of daily rainfall records over several decades to train and evaluate each model. The effectiveness of the models was measured using various metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE), which shed light on their precision and dependability. Initial results showed that LSTM surpassed both ARIMA and SARIMA in performance, demonstrating superior accuracy and affirming its strength as a robust deep learning tool for capturing intricate patterns. ARIMA’s performance was notably weaker, highlighting its difficulties with data featuring inherent seasonal tendencies. Although SARIMA made improvements over ARIMA’s flaws by integrating seasonality, it still did not reach the accuracy levels of LSTM. In conclusion, the findings from this project supported the use of LSTM models for weather forecasting, especially in predicting rainfall. The outcomes emphasized the necessity of selecting suitable modeling techniques based on the specific characteristics of the data and the requirements of the forecasting task. Looking ahead, future research might investigate hybrid models that merge the strengths of LSTM with traditional models like SARIMA, potentially creating even more powerful forecasting tools. This work contributes to the ongoing efforts in the meteorological field to enhance prediction accuracy through sophisticated computational methods, helping to better prepare for and respond to weather-related events.

Library of Congress Subject Headings

Rain and rainfall--Forecasting--Data processing; Machine learning

Publication Date

12-2024

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Graduate Programs & Research

Advisor

Sanjay Modak

Advisor/Committee Member

Ioannis Karamitsos

Campus

RIT Dubai

Plan Codes

PROFST-MS

Download

Share

COinS