Towards Smarter Energy Planning: District-Based Natural Gas Consumption Forecasting in Istanbul with BI-GRU

With the increasing demand for energy, accurate forecasting of natural gas consumption in large cities has gained strategic importance. This paper compares the performance of various forecasting models using natural gas consumption data of districts in Istanbul, the most populous city in Türkiye. Using a time series analysis approach, the deep learning models Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BI-LSTM), Bidirectional GRU (BI-GRU) and Recurrent Neural Network (RNN) are compared with machine learning models Linear Regression (LR), Ridge, Support Vector Regression (SVR), Random Forest (RF) and Gradient Boosting algorithms. Model performance was evaluated with R² and Mean Absolute Percentage Error (MAPE) metrics. According to the results, the BI-GRU model provided the highest accuracy with a MAPE of 12.98% and an R² value of 0.8902. LSTM, BI-LSTM and RNN models also showed high success. In contrast, linear regression-based models performed poorly with low R² and high MAPE values. Among the machine learning methods, only the Random Forest model gave a strong result with an R² of 0.9063, while the MAPE value remained high. As a result, deep learning models, especially BI-GRU, better capture short-term consumption fluctuations and provide better forecasts compared to classical models.