The German industry has to reduce its CO2 emissions – a fact virtually everyone agrees on. But how?

The factories of Bosch are also facing this question. They have already reduced their CO2 emissions, but want to reduce them even more.

A key element in reducing CO2 emissions is the optimized use of power, cooling, and heat – it can contribute to reducing energy consumption and saving costs.

This optimization requires a comprehensive model of energy consumption, generation and flows depending on control interventions, and external factors such as weather. The HeatCast project has set itself the goal of developing such a forecasting model based on the data provided by Bosch. This is done using temporal fusion transformers (TFT), which are particularly suited to solve multi-horizon time series problems, such as forecasting heat, cold water, and power consumption.

The project is interesting to:

Industrial enterprises, utility companies, grid operators


  • Developing a forecast for heat and cooling consumption
  • Developing a forecast for power consumption
  • Comparing temporal fusion transformers and extreme learning machines
  • Visualization in the EnergyPilot energy management system


  • Temporal fusion transformers for forecasting the time series
  • Extreme learning machines for forecasting the time series for comparison
  • EnergyPilot, IEE’s energy management system

Detailed project description

The HeatCast project investigates whether temporal fusion transformers (TFT) are suitable for calculating power and hot/cold water consumption forecasts. Forecasts are prepared using extreme learning machines (ELMs) and TFT, and the two methods are compared with each other.

TFT is based on the principle of transformer networks. Known future information, such as day of the week, time, sun position, and production targets, can be included in the forecast to better recognize temporal patterns. Alongside heterogeneous utilization of the input data, TFT also offer the possibility of preparing multi-horizon and probabilistic forecasts.

Extreme learning machines (ELMs) are based on an artificial neural network with only one hidden layer, where the number of neurons is selected to be very high. Whereas the links between the input layer and the hidden layer are assigned random weights, only the weights between the hidden layer and the output layer are included in the training. This creates a simple linear regression problem that can be quickly analyzed and solved. The ELM is trained in an environment that is also used for preparing forecasts for productive operation.

Both methods demonstrate to be capable of forecasting the power as well as hot/cold water consumption, but they differ in terms of forecasting quality. The forecast enables the Bosch production facility to control its own processes, thereby saving power, reducing cooling water consumption, and consequently reducing CO2 emissions.

Integration into EnergyPilot, the energy management system of Fraunhofer IEE, offers the possibility of adapting the control of power consumption and hot/cold water consumption. Another development option is the integration of forecasts for the facility’s PV system into operational process control to maximize the consumption of power generated by the PV system.

Project partners

Business Unit Energy Meteorology Information Systems, Fraunhofer IEE

You can find more information about the business unit here.

Project participants: Dr. Klara Reder, Alexander Dreher, Thomas Kanefendt, Jonas Koch, Malte Lehna


Within Bosch, HeatCast is part of the Zero Emission project. The project is aimed at making a large production site of Bosch in Thuringia carbon-neutral. The Eisenach production site employs 1,700 staff members. The site manufactures products for the large Bosch divisions Powertrain Solutions, Chassis Systems Control, and Automotive Aftermarket.

Point of contact: Wilma Weps

Project period

April 2021 – October 2021

Dominik Jost

Verbrauchsmodellierung- und Prognose

Fraunhofer IEE

+49 (0) 561 7294-467

Share this Spotlight with your network.

GRAF-CoWi - Combination of wind power forecasts
cognition2hydrogenforceCognition2H2Force - Efficient hydrogen production