In modern energy systems, the high volatility of electricity prices makes electricity price forecasting (EPF) a fundamental pillar for effective energy supply and inventory operations. However, accurate forecasting is becoming increasingly difficult due to unstable renewable energy, fluctuating demand, and an overwhelming number of market variables. Thus, to effectively mitigate risks and manage energy supply, forecasting models must be both accurate and capable of quantifying price uncertainty. To this end, we propose Reinforcement Learning-driven Time-series Dense Encoder (RL-TDE), a probabilistic forecasting framework that integrates Reinforcement Learning-based feature selection with an efficient temporal encoding scheme. Specifically, a Double Deep Q-Network (DDQN) agent dynamically identifies informative feature subsets to reduce redundancy and enhance generalization. These features are then processed by a lightweight Time-series Dense Encoder optimized for fast inference and quantile forecasting. Extensive evaluations on real-world datasets demonstrate that RL-TDE significantly outperforms traditional and deep learning baselines in both accuracy and interval reliability. Finally, we validate the framework’s operational value by applying it to an energy storage system. A multi-turn strategy shows that our forecasts drive profitable inventory operations and arbitrage, confirming the model’s utility for operational decision-making in energy supply chains.