Abstract:
Accurate temperature forecasting is vital for agriculture, disaster management, and climate resilience, particularly
in semi-arid regions like Machakos County, Kenya. Traditional forecasting models, such as the Auto Regressive Integrated
Moving Average (ARIMA), have been widely used for both short- and long-term temperature predictions. However, with
advancements in machine learning, there is a growing need to evaluate how modern methods compare to traditional approaches.
This research focused on comparing the predictive accuracy of ARIMA and the Long Short-Term Memory (LSTM) model, a
deep learning algorithm that uses a gating mechanism to capture non-linear patterns in data. The study used Mean Absolute
Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Diebold-Mariano
statistical test to evaluate and compare the performance of models. Temperature data was obtained from the National
Aeronautics and Space Administration’s Prediction of Worldwide Energy Resource (NASA POWER) database, using GPS
coordinates to retrieve location-specific data for Machakos County. To ensure robust analysis, Python libraries such as
Statsmodels, Pandas, TensorFlow, NumPy, Keras, and Matplotlib were used for data processing, fitting of models, and
visualization of results. Findings showed that LSTM performed better in long-term predictions, achieving higher accuracy for
30-day forecasts, while both models performed significantly equivalently in the short term of 7 days. These findings highlight
the complementary strengths of traditional and deep learning models in addressing different forecasting needs.
