Real time weather forcasting with conditional CNN and TCN-BiLSTM Ensamble at Manokwari

Abstract

Short-term weather forecasting is fundamentally critical for disaster mitigation in dynamic tropical maritime regions. However, conventional numerical approaches suffer from high computational latency, and spatial deep learning models frequently experience severe performance degradation during nocturnal conditions due to the absence of illumination. This study aims to develop an adaptive, real-time multimodal weather nowcasting system that effectively prevents nocturnal predictive failure through a dynamic conditional ensemble architecture. The proposed framework integrates a Convolutional Neural Network (CNN) to extract optical features from a dataset of 2,515 localized sky images with a Temporal Convolutional Network and Bidirectional Long Short-Term Memory (TCN-BiLSTM) pipeline to process 15,111 corresponding meteorological time-series records from BMKG. To address visual ambiguity, the system strictly employs a day-night gating mechanism, deactivating the CNN at night to rely solely on thermodynamic data. Finally, the optimized model was deployed via TensorFlow.js for decentralized client-side browser inference. Experimental evaluations explicitly demonstrate that the conditional ensemble significantly outperformed all standalone models, achieving a peak accuracy of 92.49% and a Macro F1-score of 0.913 while successfully preserving a robust recall rate for precipitation events. Furthermore, the browser-based deployment completely eliminated server transmission bottlenecks, achieving sub-second warm-start inference latency across heterogeneous consumer devices. Ultimately, the conditional day-night modality gating mechanism effectively mitigates nocturnal visual degradation, proving that implementing this integrated architecture as a client-side web application is highly feasible for delivering instantaneous public weather warnings.

GS Cited Analysis