Integrated energy systems that combine electricity, heating, cooling, and other energy vectors face mounting challenges in accurately forecasting demand. Climate variability, renewable energy volatility, and complex interdependencies between different load types have made traditional forecasting methods insufficient. A new machine learning approach addresses these limitations through an advanced neural network architecture.
The proposed MTL-TCN-Transformer model combines three complementary techniques. Temporal Convolutional Networks capture long-range dependencies in time-series data, while Transformer layers establish global correlations across different energy loads. Multi-task learning enables the model to simultaneously forecast multiple load types while sharing learned features, improving prediction accuracy and generalization.
A key innovation lies in feature engineering using maximum information coefficient methods to identify and construct nine correlation features between loads—including load ratios and differences. This approach helps the model understand how electricity, thermal, and other energy demands interact and influence one another.
Testing across four seasons showed average Mean Absolute Percentage Error (MAPE) reductions of 67.95% for electrical loads compared to conventional methods. The model maintains robust performance despite seasonal variations and climate-driven volatility that traditionally degrade forecasting accuracy.
For grid operators and energy planners, improved load forecasting directly translates to better dispatch decisions, reduced reserve margins, and more efficient integration of distributed and renewable resources. The model's ability to forecast multiple energy vectors simultaneously supports the growing trend toward sector coupling—where electricity, heating, and transportation systems are increasingly integrated.
Future work will focus on expanding cross-regional generalization capabilities, enabling the model to transfer learning across different climate zones and grid configurations. This advancement positions utilities and grid operators to better manage the transition toward climate-resilient, flexible energy systems.
