Electronic Theses and Dissertation

Danau Maninjau mengalami peristiwa upwelling secara berkala yang mengganggu kualitas air, merusak stok ikan, dan menimbulkan tantangan sosial ekonomi bagi masyarakat sekitar. Penelitian ini bertujuan untuk meningkatkan akurasi prediksi upwelling dengan mengintegrasikan pemodelan deret waktu Vector Autoregressive (VAR) dengan klasifikasi Support Vector Machine (SVM). Dataset lima tahun (2020-2024) dari variabel iklim harian suhu permukaan, curah hujan, dan kecepatan angin dikumpulkan dari National Aeronautics and Space Administration (NASA). Stasioneritas data dikonfirmasi dengan menggunakan transformasi Box-Cox dan uji Augmented Dickey-Fuller, sementara analisis Kausalitas Granger menunjukkan hubungan dua arah di antara variabel-variabel tersebut. Model peramalan yang optimal, VAR (17), dipilih berdasarkan Akaike Information Criterion (AIC), yang memastikan residual memenuhi kriteria white noise. Pengelompokan K-means kemudian memberi label pada hari-hari upwelling potensial, dan label-label ini digunakan untuk melatih pengklasifikasi SVM. Dasbor interaktif dikembangkan dengan menggunakan Python dan Streamlit untuk memfasilitasi prakiraan waktu nyata dan hasil klasifikasi. Model VAR (17) menghasilkan prakiraan yang sangat akurat, yang tercermin dari metrik kesalahan yang minimal (misalnya, RMSE

Lake Maninjau experiences periodic upwelling events that disrupt water quality, harm fish stocks, and pose socioeconomic challenges to surrounding communities. This study aimed to enhance upwelling prediction accuracy by integrating Vector Autoregressive (VAR) time series modelling with Support Vector Machine (SVM) classification. A five-year dataset (2020–2024) of daily climate variables surface temperature, precipitation, and wind speed was collected from NASA. Data stationarity was confirmed using Box-Cox transformations and Augmented Dickey-Fuller tests, while Granger Causality analysis revealed bidirectional relationships among the variables. The optimal forecasting model, VAR(17), was selected based on the Akaike Information Criterion (AIC), ensuring residuals met white-noise criteria. K-means clustering then labelled potential upwelling days, and these labels were employed to train SVM classifiers. An interactive dashboard was developed using Python and Streamlit to facilitate real-time forecasts and classification outputs. The VAR(17) model produced highly accurate forecasts, reflected by minimal error metrics (e.g., RMSE < 0.60). SVM classification of potential upwelling events achieved strong performance, consistently attaining F1-scores above 0.95. By merging time series forecasts with event classification, the hybrid VAR–SVM framework outperformed single-method approaches in identifying and predicting upwelling episodes. This integrated modelling strategy effectively addresses the complexity of upwelling in Lake Maninjau, enabling timely decision-making for fisheries management and local tourism stakeholders.