Predictive Modeling for Natural Gas Prices in Romania Using Time Series Data and Average Temperature

The correct prediction of natural gas prices is an essential tool for energy market decision strategies. The research on deep learning models shows that LSTM models outperform traditional econometric models because they effectively detect the non linear patterns and price instabilities in commodities. The literature review demonstrates that incorporating external variables including weather and economic indicators may lead to better forecasting results. This research implements an LSTM forecasting model to predict Romanian day-ahead natural gas prices by including daily average temperatures as external variables. This study employed partial autocorrelation analysis to determine appropriate lag lengths and RobustScaler to handle outliers for stable model training. The research methodology employs daily data from 2019 to 2024 spitted into training and testing data. The model’s performance was evaluated with Mean Squared Error (MSE), Coefficient of Determination (R²), and Mean Absolute Percentage Error (MAPE). The results showcase a high predictive accuracy, with an MSE of 0.002839, an R² of 0.8886, and a MAPE of 4.10% on the test data. These results shows how LSTM with exogenous variables effectively predict market fluctuations when combined together and how temperature acts as a key factor for short and mid range market prediction. The energy sector benefit substantially from accurate forecasts because they enable better risk management, strategic planning and regulatory policy decision-making. The study enhances recent data-driven energy market forecasting research through its implementation of advanced neural architectures and external factors which establishes a base for future investigations into hybrid methods and additional exogenous variables that represent changing market conditions.