Accurate short-term electricity price forecasting is essential for effective demand-responsive energy management. Reinforcement learning schedulers use anticipatory signals to optimize appliance usage and storage dispatch. This research studies the forecasting accuracy and training dynamics of a Multilayer Perceptron (MLP) model using household demand and locational marginal price (LMP) data. The proposed three-layer feedforward architecture integrates temporal and contextual features, such as lagged demand, price histories, and calendar variables. It is trained by backpropagation with the Adam optimizer and early stopping. Convergence analysis indicates consistent learning behaviour with a minimal generalization gap, ensuring reliable predictions under new test conditions. Compared against persistence, ARIMA, GRU, and LSTM baselines, the MLP outperforms them with an RMSE of 0.0081 $/kWh, MAPE of 3.46%, and R² of 0.932, while retaining inference times below 0.3 ms per sample. Residual error analysis demonstrates the absence of systematic bias and less variance as compared to recurrent models. These findings establish the MLP as a computationally efficient and statistically reliable forecaster that enhances reinforcement learning-based scheduling and saves costs in smart-home demand response applications.

Keywords

Short time electricity price forecast, multi-layer perceptron (MLP), smart home energy management, reinforcement learning scheduling, demand response optimization