XIII-04: Forecasting DHC demand by AI by Moritz Bitterling

Welcome to my presentation about evaluating different artificial neural network approaches for forecasting heat demand in district heating networks. My name is Moritz Bitterling and I'm a member of the research group Climate Neutral Industrial Processes and High Temperature Storage at the Fraunhofer Institute for Solar Energy Systems in Freiburg in Germany.

So first, about our motivation and the background. Why did we want to apply neural networks to a district heating network? We had access to a combined residential and industrial district heating network in South Germany in Lebenau. You can see that on the map on the right side. And together with the owner and the operator of the network, our goal was to find a simple and general methodology to improve improved operation strategy of this district heating network, which he can directly apply. And we choose the approach with the artificial neural networks because it's relatively simple and it can easily be added to the existing system. And with this approach we wanted to quantify the optimization potentials that we see or that we get for this network. And the goal is to enable a real time operation of the INN system to support the operators of this district heating networks in taking the decisions which would lead to cost savings and emission reduction of this district heating network. And of course, the aim is to transfer this methodology to other networks as well. The outline of my presentation will be that I will shortly talk about the district heating network that we worked on. Then introduce the different ANN architectures that we used and show how the training procedure worked. And then I will present the results performance evaluation of those trained ann's.

This is the district heating network in Lebanon. The network layout the annual heat consumption is around 15 GWh and the network consists of several heat producers, for example, Combined Heat and power block, wood chip boilers, a diaper incinerator and also gas and oil boilers and consumers. In the network are around 40 residential and office buildings and laundry, several workshops, wood and hay drying and also two greenhouses.

The data that we have from the DHN starts at the beginning of last year. We have more than 100 measurement points and we measure supply and return temperatures, flow rates, power meter readings and ambient temperature at different locations. As an example, in the top right graph you can see power consumptions for different buildings in the network. And in addition to the measurement data from the network, we.

Collect and download weather forecasts from MeteorBlue for this location twice a day.

And now about the different kinds of neural networks that we trained on that data. The first type of networks was convolutional neural network. In this kind of network, two dimensional filters are applied to the Data in convolutional layers. And this kind of network is mainly used in image processing tasks. And the second type of network is a long short term memory network. It's a type of recurrent neural network which process time series step by step. And by doing that, they maintain internal state from time step to time step. The main use of this kind of networks is to process time series, for example speech recognition. And then with those two kinds of networks, there are different ways of predicting the output. One way is called single shot.

In this kind of application, the output is for different time steps, is produced in one step. And the other way of applying those networks is called autoregressive, where the output is predicted step by step. So always one data point in the future is predicted. And while we were training those networks, we varied hyper parameters to optimize the prediction quality of those networks.

In more detail, this is how it looks. When we use a network to produce a certain output from input. We have a known past and partly known future, which is the weather forecast. And we want to predict the unknown future, which is the heat demand. In the this case.

To create a training data set for an ann, it's preferred to have more than one year of data, because then the behavior of the DHN all seasons can be captured. And we train the ANN on the whole data set. And then we can use the trained ANN to predict in real time the heat demand of the dhn. When we do the training, we do not use the or we did not use the whole data set for training, but we need part of the data set for validation data in order to avoid overtraining of the ann. In our case, we used March and July and December as a validation data set.

Here you can see different predictions of a trained annual.

In this figure, on the x axis there is the time and on the Y axis there is the heat demand. Blue is the real heat demand of the dhn and in orange there is the prediction of the neural network. And the dashed line in the center of the plot is the moment of the prediction. So you can see that for example, in the second plot the trend is captured very well. But in the first and the last plot, stronger fluctuations in the heat demand are not so well described. And these are all predictions made by a single shot cnn. But the interesting result that we saw is that all optimized network architectures that we analyzed, they showed almost the same prediction quality.

This is a more general statistical evaluation, not only for single points in time, but what we did here is that for the Training data set for every full hour we made a 12 hour forecast and then calculated the difference between the forecast and the real heat demand. And what you can see here, the uncertainty intervals are shown. It means that for example, for the yellow band, 90% of the differences between forecast and real heat demand are within this band and 80% of the differences are within the green band and so on. And what is interesting is that the prediction error is almost constant over the 12 hour prediction interval. The mean 80% confidence interval over the 12 hours is around plus minus 200 kilowatts and the mean absolute error in that time is 135 kilowatts. For a better interpretation of those absolute values, we scaled them to the maximum heat demand and then we get a mean absolute error of around 5% and and 80% confidence interval of plus minus 8% deviation.

So here on the left side you can see what I just showed. The mean absolute error of the predictions in our data set is around 5% and the 80% confidence intervals of the deviations are plus minus 8%. If you now only predict the heat demand of office and residential buildings, the results between become slightly better. So this means that the behavior of the industrial buildings in the DHN is a bit less predictable. And if we compare the prediction quality with an older approach that we have here, which uses a linear system identification combined with a more simple ann, the results of the single shot CNN are almost twice as good as.

So to conclude we see a similar performance of convolutional neural networks and long short term memory networks and they both outperform older AI approaches. The heat demand of the DHN can be predicted with reasonable accuracy. So for a prediction of 12 hours in advance with 1 hour resolution, we have a probability of 80% that the prediction error is less than 8% of the maximum heat demand of the network. Trends in the data can be captured well, but short peaks and fluctuations are harder to catch. So we see that AI forecasting can be implemented in this DHN at low cost and that the economic benefit compared to this cost is. Is very promising.

Thank you very much for listening to my presentation. If you have any questions, please get in contact with us by phone or email. Our contact details are given here on slide. Finally, I would like to express our gratitude to the IEA Technology Collaboration Program on District Heating and Cooling for financial support and for the feedback and special thanks also to our project partner Stiftungwibenau for the access to the heating network and for the many discussion rounds. And I would like to point out to you that the final report of our project will be published on the IAEA website. In this report, there is more detailed information about what we did. And you can also see the link to the website and to the report here on this slide. Thank you very much.