Global district heating forecasting models

The key to better forecasting and more efficient systems

09.05.2025 – ca. 9 Min. Lesezeit – Zurück zur Startseite – Alle Blog-Artikel

Accurately predicting heat demand in district heating systems is a technical challenge and a necessity for efficient and sustainable energy supply.

Advanced time series models are increasingly being used to meet this challenge. However, despite advances in modeling, a key question remains: How well can these models be applied to new, previously unknown district heating systems?

This question is particularly relevant for new district heating systems for which limited historical data is available, a situation often referred to as the “Cold-Start-Problem. In such cases, models must be able to learn from the data of several existing systems to make accurate predictions for systems without extensive historical data.

The importance of generalisation for prediction accuracy

The generalisability of a prediction model describes how well it can apply learned patterns to new, unknown data.

This is particularly important in district heating forecasting, as each system has specific characteristics and seasonal variations. A model must therefore not only make accurate predictions for the data of the systems on which it has been trained, but also apply to new, previously unknown systems – even if limited historical data is available. This is a significant challenge, as the diversity and complexity of district heating systems require accurate and reliable modelling.

Modern models: The Temporal Fusion Transformer (TFT) and TimesFM

In recent years, transformer-based models such as the Temporal Fusion Transformer (TFT) and the Time Series Foundational Model (TimesFM) have set new standards. These models offer an extended generalisation capability, particularly advantageous for heterogeneous district heating systems.

A particular feature of these models is their ability to operate as global models instead of traditional local models that are only trained on data from a single system.

Global models vs. Local models

Traditionally, local models forecast heat demand in district heating systems. This means that a separate model is trained for each district heating system, based only on the historical data of that specific system. Such a model cannot recognise and exploit synergies between different systems, which is particularly problematic when little historical data is available for a new system (e.g., in the case of the Cold-Start-Problem). In addition, the maintenance effort increases as a separate model must be maintained for each system.

Global models such as TFT and TimesFM take a different approach: instead of training a separate model for each system, these models are trained on data from multiple systems simultaneously. This means the models can benefit from the synergies and common patterns between different systems. Suppose different district heating systems have similar seasonal variations or similar patterns of heat demand. The model can use this information to make more accurate predictions, even for systems with limited historical data.

Advantages of global modeling

The use of global models has some significant advantages:

Improved generalisation
Global models can not only learn patterns within a single system, but also recognise broader relationships between different systems. This enables them to make accurate predictions even when data for a specific system is missing or limited.
Using synergies between systems
Different district heating systems may show similar patterns that a global model can recognise. Such a model can, for example, simultaneously take into account seasonal influences or weather-related fluctuations in the heat demand of different systems, thus increasing the accuracy of the forecast.
Reduction in the effort required for model maintenance
Instead of developing and maintaining a separate model for each system, a single global model works for several systems simultaneously. This significantly reduces the maintenance effort, especially when new systems are added.
Better handling of the Cold-Start-Problem
The cold-start problem, where limited historical data is available for new systems, is better addressed by global models. By accessing data from multiple systems, global models can recognise patterns and correlations that would be difficult for a single system to identify.

TFT and TimesFM as global models

The Temporal Fusion Transformer (TFT) is an example of a global model. It uses its attention mechanisms to identify relevant features across multiple time series. It takes into account not only the historical data of a system, but also external variables such as weather conditions or system size, which are crucial for predicting heat demand. This allows the TFT to exploit synergies between different district heating systems, improving the accuracy of predictions even in new, less well-documented systems.

The TimesFM model can also be used as a global model. However, it is a so-called univariate model, i.e., it does not consider external covariates, but focuses exclusively on the historical time series of a system, i.e., exclusively on the evolution of heat demand. However, TimesFM has the advantage of being based on extensive pre-training, which makes it particularly powerful when applied to heterogeneous datasets. It also offers high scalability and helps predict univariate time series in situations with little historical data.

These models demonstrate how global model architectures can help reduce complexity and effort in the energy sector while increasing forecast accuracy and flexibility, especially in scenarios with multiple heterogeneous district heating systems.

Evaluation of generalisation capability

As part of my Master’s thesis, the generalisation capability of these modern models was tested using real data from different district heating systems. Both internal generalisation (performance on new data from systems included in the training set) and external generalisation (performance on new, unseen systems) were evaluated.

The results of the internal generalisation tests showed that the TFT model, as a global multivariate model, can significantly outperform the previous local models. It should be emphasised that the TFT model can integrate many additional data sources, leading to improved prediction accuracy. The TimesFM model, which can rely on extensive pre-training, performed worse. An explanation could be that the TimesFM, as a univariate model, does not consider external covariates such as weather data or system variables. This additional information, which the TFT model can incorporate due to its multivariate architecture, allows for a more accurate capture of heat demand patterns and leads to more reliable predictions.

The advantage of the global models was evident in the external generalisation test, which simulated the Cold-Start-Problem. With only four days of historical data, the TFT and TimesFM models could make significantly more accurate predictions than a traditional local model based on the limited data from a single system. Again, the TFT model was superior, providing slightly more accurate predictions than the TimesFM model.

The evaluation results confirm that global multivariate models, such as the Temporal Fusion Transformer (TFT), offer significant advantages in prediction accuracy, especially in complex and heterogeneous district heating systems. With the ability to exploit synergies between different systems, these models significantly improve over traditional local models. In addition, they show their strength in dealing with Cold-Start-Problems, as they can provide reliable predictions even with limited historical data.

Author

, Working student at scieneers GmbH
zoe.zantow@scieneers.de