Nowadays, renewable energy sources are gaining importance, yet global energy demand is primarily met by burning fossil fuels. Fluctuations in fossil fuel availability, driven by geopolitical tensions, supply-demand changes, and natural disasters, can lead to sudden energy price spikes or supply shortages, adversely affecting the global economy. Despite its negative impact on carbon emissions and climate change, Heating Oil (HO) offers advantages over other fossil fuels in efficiency, reliability, and availability. Accurate time series prediction models for HO are crucial for stakeholders. This study proposes a novel hybrid model, integrating the Chaotic Adaptive Fitness-Distance Balance-based Stochastic Fractal Search (AFDB-SFS) algorithm with a Bidirectional Long-Short Term Memory (Bi-LSTM) network, for HO close price prediction. The dataset comprises daily observations of five financial time series (close, open, high, low, and volume) over 4260 trading days, yielding a total of 21,300 data points (4260 days x 5 variables). During the feature extraction stage, financial signal processing methods such as Demand Concentration Curve (DCC) and traditional technical indicators are utilized. A total of 189 features are extracted at appropriate intervals for each indicator. Due to the large number of features, the AFDB-SFS algorithm then efficiently identifies the most compatible feature subsets, optimizing the Bi-LSTM model based on three criteria: maximizing R2, minimizing RMSE, and minimizing feature count. Experimental results demonstrate the proposed hybrid model's superior performance, achieving high accuracy (R2 of 0.9959 and RMSE of 0.0364), outperforming contemporary models in the literature. Furthermore, the model is successfully implemented on the Jetson Orin Nano Developer Platform, enabling real-time, high-frequency HO price predictions with ultra-low latency (1.01 ms for Bi-LSTM), showcasing its practical utility for edge computing applications in commodity markets.