This study presents a resource-aware deep learning pipeline for next-step multivariate time series forecasting, featur- ing an attentional bidirectional long short-term memory archi- tecture. The proposed model is designed to capture both forward and backward temporal dependencies while dynamically focusing on the most salient time steps using a Bahdanau-style attention mechanism. We first evaluate the method on the widely used Jena Climate dataset, then extend the study to a larger real-world me- teorological dataset from California. This combination provides both a standard benchmark and a more challenging real-world test case, where the model demonstrates its superior predictive ac- curacy compared to state-of-the-art models, including CNN-RNN, CNN-LSTM, and Stacked-LSTM baselines. To ensure practical applicability in real-world, resource-constrained environments, the entire model is optimized and deployed on a NanoPi Neo Plus2 board—an ARM-based 64-bit single-board computer with limited computational resources. Our implementation leverages lightweight inference techniques and efficient model quantization to enable on-device prediction without cloud connectivity. The resulting system achieves competitive forecasting performance with minimal latency and power consumption, showcasing the feasibility of edge-AI solutions for environmental monitoring and smart sensing applications. Both quantitative and qualitative analyses confirm the effectiveness and interpretability of the proposed approach.The source code for this project is publicly available at: https://github.com/NavidH95/attentional-bilstm-for- edge-forecasting/tree/main.

Keywords