Forecasting heat consumption

We are working on forecasting heat consumption ― and we invite you to take a look into our R&D kitchen.

In particular, our talented intern Minh Nguyen worked on finding a good model for data we use in the TUNES project. In this post, we’ll cover the following points:

1. Why is forecasting heat consumption useful?
2. What was our approach, and how do results look?
3. What is tricky about forecasting domestic heating consumption data?

Let’s begin!

We’ve been working on a few projects around the topic of smart heating. For us, smart heating is storage optimization. Heat can be buffered for later use, which means creating the heat is a flexible energy process. And flexibility is where we come in 🙂

Heat can be buffered in a well-insulated building, but also in a “heat battery” ― basically a hot water tank (which also needs very good insulation). We can create the heat with electricity (using heat pumps or other means). The times we use electricity can have meaning, as we:

• pay a low energy price when we choose the right moments
• we use our locally generated solar power
• we help the grid operator maintain stable operations by reacting to local problems

Of course, planning involves the future, so we need to know what heat demand might be coming. Will the occupants be at home, when outside temperatures are low? Will they use the shower a lot today, because they all play sports on Saturday?

For our planning, to deliver workable services, this matters.

First of all, what makes this study special? It focused on residential buildings (built before 2000). Usually a fixed-point forecast is studied, but we are interested in multi-point forecasts (what if we predict the same hour of the day while we come closer to it? – like weather forecasts, we should get better the closer we are in time) and probabilistic forecasting (we forecast a consumption quantity, but how sure are we?)

Minh first studied data of the apartments of the TUNES project. What are the patterns? Do they happen in a very repeated fashion? similar across apartments?

Taking into account flow rate of the central heating, in- and outflow temperature of the heat pumps, the following picture emerges – surprise, people are different when we look so closely:

To build a forecasting model, you need to decide what algorithm to use, but also (quite importantly) what information to feed into the model.

As input parameters, Minh looked at:

• Hour of the day
• Ambient (outside) temperature
• Heat consumption one hour ago, 24 ago, 48 ago

Historical consumption matters most, ambient temperature helps.

Three algorithms were tried, and the one that performed best was Light Gradient Boosting (LightGBM) with Quantile Regression.

Here are visualizations from the forecasting performance using our internal tool. The black line is consumption, the blue line is the forecast, and you can see the uncertainty window around the forecast:

Here is another visualization from our tool: Accuracy per hour of the day (basically an accuracy heatmap). We can be super-accurate in the night (dark green), because heating does not really happen at night, when people sleep. Heating consumption is always zero. The morning seems to be the most difficult time to get exactly right.

But how right do we have to be to build a well-working heat buffering system? I would say we have to get the overall amount right, more than the exact time it will be used. If we set a lower boundary for the buffer correctly, we should be fine

First, forecasting the heating behavior of single domestic households is not easy. Forecasting a thousand households is much easier, due to statistics. This makes smart energy solutions in the domestic sector a bit tricky to scale.

Second, this data has many zeroes, as households often do not heat when they are asleep or away from home for some reason. Zeroes are numbers, but they are indicating absence of activity. This makes it tricky for models to capture patterns and model them well mathematically.

Finally, resampling data (i.e. by the hour, as we did here) in combination with zero values, can make it hard for ML models to use weather data (like ambient temperature, in our case). For example, sometimes the outside temperature can fluctuate from really high to really low, but the demand can still be zero for the whole time because the occupant is not home. The ML model does not know if they are home (for privacy reasons), so it will have reduced modeling accuracy with respect to the temperature. Minh therefore saw that mathematically, the influence of outside temperature on the forecast accuracy was diminished, even though in practice it plays a greater role.

Well, we hope this view into our kitchen was interesting. We will use these outcomes in the TUNES project, but we already have some other collaborations lined up where we also need to forecast heat consumption (of apartments, large buildings or even SMEs).

If you are interested into heating forecasting or planning, get in touch. You can also reach Minh directly, who will go on to study at LMU Munich (where he will go on do great things I’m sure!).