Browsing by Author "Kabue, Ceasar Waweru"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Weather Forecasting using Radial Basis Function Network
    (Kenyatta University, 2025-05) Kabue, Ceasar Waweru
    Weather forecasting plays a vital role across sectors such as agriculture, disaster management, transportation, and urban planning. While traditional models like Numerical Weather Prediction (NWP) and Autoregressive Integrated Moving Average (ARIMA) have been instrumental, they often fall short in handling the nonlinear and chaotic nature of atmospheric systems, especially in long-term predictions. This study addresses these limitations by applying and evaluating the performance of machine learning models—Radial Basis Function Networks (RBFNs), Convolutional Neural Networks (CNNs), and Long Short-Term Memory networks (LSTMs)—using a decade-long dataset (2013–2023) obtained from the Kenya Meteorological Department, encompassing temperature, humidity, rainfall, sea-level pressure, and windspeed. Data preprocessing involved K-Nearest Neighbours (KNN) imputation and Min-Max normalization. Models were developed using TensorFlow in Python, optimized through grid search, and deployed via Docker containers on Google Cloud Platform to simulate operational forecasting conditions. Comparative analysis showed that RBFNs outperformed CNNs and LSTMs across all variables, achieving the lowest Root Mean Squared Error (e.g., 0.239°C for temperature versus 5.975°C and 8.701°C for CNN and LSTM, respectively) and the fastest training time, making them highly suitable for real-time forecasting in resource-constrained environments. Hybrid models combining RBFNs with CNNs and LSTMs showed improved accuracy, particularly for complex variables like rainfall. These findings suggest that RBFNs, either standalone or in hybrid configurations, provide an efficient, scalable, and accurate alternative for operational weather forecasting. The study concludes by recommending further exploration of hybrid model integration and the incorporation of satellite imagery to enhance spatial resolution.
  • Thumbnail Image
    Item
    Weather Forecasting using Radial Basis Function Network
    (TIJER – International Research Journal, 2024-10) Kabue, Ceasar Waweru; Matheka, Abraham Mutua
    Weather forecasting has been a critical component in various industries such as agriculture, disaster management, transportation, and urban planning. Accurate weather predictions have helped minimize disruptions, enhance decision-making, and reduce economic losses. Traditional forecasting models, including Numerical Weather Prediction (NWP) and Autoregressive Integrated Moving Average (ARIMA), proved effective but faced challenges due to the nonlinear and chaotic nature of weather systems. Minor errors in the initial conditions of these models resulted in substantial inaccuracies, especially for long-term forecasts. This phenomenon, commonly referred to as the "butterfly effect," highlighted the inherent limitations of traditional models in capturing the complexity of atmospheric systems. In response to these challenges, machine learning emerged as a promising alternative, offering the ability to manage vast amounts of complex, nonlinear data. Machine learning models, particularly Artificial Neural Networks (ANNs), demonstrated considerable success in short-term and mid-term weather forecasting. These models identified and generalized patterns in meteorological data that were not apparent through traditional methods. Among ANNs, Radial Basis Function Networks (RBFNs) showed potential due to their efficient handling of time-series data, fast training times, and ability to model nonlinear relationships with noisy inputs. This study explored the performance of RBFNs in weather forecasting and compared their effectiveness to advanced deep learning models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs). While CNNs were highly effective in extracting spatial features from satellite data and other weather imagery, LSTMs excelled in learning temporal dependencies, making them suitable for long-term weather forecasting tasks. However, both CNNs and LSTMs proved computationally expensive, requiring large datasets, extensive training times, and significant computational resources, which limited their application in real-time forecasting, especially in resource-constrained environments. The primary objective of this research was to evaluate the comparative performance of RBFNs, CNNs, and LSTMs in weather forecasting, focusing on accuracy, computational efficiency, and training time. Using historical weather data from the Kenya Meteorological Department, spanning from 2013 to 2023, the study assessed the predictive power of these models for key meteorological variables such as temperature, humidity, windspeed, sea-level pressure, and rainfall. The results indicated that RBFNs consistently outperformed CNNs and LSTMs, particularly in terms of computational efficiency and accuracy, making them a more viable option for real-time applications where speed and resource efficiency were critical. Additionally, this research highlighted the potential of hybrid models that combined RBFNs, CNNs, and LSTMs to leverage the strengths of each architecture. While RBFNs offered rapid real-time predictions, CNNs provided the spatial accuracy required for analyzing satellite imagery, and LSTMs captured long-term temporal patterns. The integration of these models significantly improved forecasting accuracy, particularly for chaotic and highly variable weather phenomena such as rainfall and windspeed. The study concluded that RBFNs were an optimal solution for weather forecasting in resource-limited environments due to their fast training times and reduced computational demands. The findings also suggested that further exploration into hybrid models could provide a more comprehensive and accurate framework for weather prediction. Future research should focus on integrating satellitebased data with ground-level observations to enhance spatial accuracy and utilizing hybrid machine learning models to combine the strengths of RBFNs, CNNs, and LSTMs. Moreover, the scalability and accessibility of these models could be improved through advanced data preprocessing techniques, model optimization, and transfer learning, ensuring their applicability in diverse geographical regions with varying levels of data availability