Photo: Empa

Ham­ster­ing cot­tages: AI for ener­gy man­age­ment in the house

The ener­gy man­age­ment in a house with a solar sys­tem is becom­ing more and more com­plex: When do I turn on the heat­ing so that it is pleas­ant­ly warm in the evening? How much elec­tric­i­ty can the hot water reser­voir absorb? Is the ener­gy still suf­fi­cient for the elec­tric car? Arti­fi­cial intel­li­gence can help: Empa researchers have devel­oped an AI con­trol sys­tem that can learn all these tasks inde­pen­dent­ly – sav­ing more than 25 per­cent energy.

How were the old days sim­ple: In the spring, when the heat­ing oil prices fell, the tanks in the base­ment were sim­ply filled to the brim. Then, until next sea­son, all the wor­ries were going on. Also for the car there was fuel at every cor­ner. Around the clock. Full refu­elling, ready to continue.

The exit from the fos­sil fuel econ­o­my makes it much hard­er for thrift fox­es. Now ener­gy prices no longer change annu­al­ly, but hourly. Solar pow­er is abun­dant at lunchtime – in the evening, the low sun hard­ly pro­vides any ener­gy, while return­ing com­muters are rapid­ly increas­ing the demand for elec­tric­i­ty in the city and the coun­try­side. The effect can be seen so clear­ly on con­sump­tion graphs that sci­en­tists have giv­en it its own name: “Duck-Curve”. When the duck rais­es its head, it becomes expen­sive for all those who now have to get electricity.

So look­ing at the clock when it comes to ener­gy would be impor­tant for elec­tric car dri­vers and home­own­ers. If you want to use the avail­able renew­able ener­gy in a cheap and envi­ron­men­tal­ly friend­ly way, you will no longer be able to rely on per­ma­nent­ly installed ther­mostats and man­u­al­ly oper­at­ed buttons.

A com­plex problem

Bratislav Sve­tozare­vic, who con­ducts research in the Urban Ener­gy Sys­tems lab­o­ra­to­ry at Empa, has rec­og­nized the prob­lem. What is need­ed is an auto­mat­ic con­trol sys­tem that ham­sters ener­gy at favourable times of the day and makes it usable for expen­sive times of the day. For exam­ple, the dri­ve bat­tery of your own car, which hangs at the charg­ing sta­tion in the garage, could serve as stor­age. But Sve­tozare­vic has to do with a com­plex prob­lem: every house is dif­fer­ent, and its inhab­i­tants are. Depend­ing on the weath­er and the sea­son, the pow­er gen­er­a­tion of the solar sys­tems as well as the need for heat­ing or cool­ing capac­i­ty changes. An opti­mal ener­gy con­trol sys­tem must there­fore learn the dai­ly rhythm of a house and its inhab­i­tants – and should also be able to react flex­i­bly dur­ing oper­a­tion, for exam­ple when a change in weath­er over­turns all calculations.

Step one: the theory

The solu­tion to such prob­lems is arti­fi­cial intel­li­gence. The Empa researcher designed an AI con­trol based on the Rein­force­ment Learn­ing prin­ci­ple. If the sys­tem acts “cor­rect­ly”, it receives a “reward”. Grad­u­al­ly, the con­troller per­fects its behav­ior in this way.

Ini­tial­ly, the con­trol was sim­u­lat­ed only on the com­put­er. The spec­i­fi­ca­tions: A cer­tain room in a build­ing had to be elec­tri­cal­ly heat­ed to the desired tem­per­a­ture and hold it. At the same time, the sys­tem had to sup­ply an elec­tric car with elec­tric­i­ty, which should be charged at least 60 per­cent at 7 a.m. in the morn­ing and go on the jour­ney. In the evening at 5 p.m., the elec­tric car returns to the charg­ing sta­tion with a resid­ual charge and can also deliv­er elec­tric­i­ty back to the house dur­ing the night hours. The con­trol sys­tem was fed with weath­er data and room tem­per­a­tures from last year and had to cope with two elec­tric­i­ty tar­iffs: expen­sive elec­tric­i­ty dur­ing the day between 8 a.m. and 8 p.m., cheap elec­tric­i­ty dur­ing the night hours.

The result was astound­ing: the self-learn­ing con­trol saved around 16 per­cent of ener­gy com­pared to a tight­ly pro­grammed solu­tion and also kept the desired room tem­per­a­ture much more pre­cise­ly in the the­o­ry experiment.

Step two: Test in the real building

Now the con­troller had to pass the test in real­i­ty. Sve­tozare­vic used NEST on the Empa cam­pus. In the DFAB House unit, the AI algo­rithm con­trolled the tem­per­a­ture of a room for a week. At the same time, the 100 kWh stor­age bat­tery in the NEST was used to sim­u­late the bat­tery of the elec­tric car. This time, the result was even clear­er: In a cool week in Feb­ru­ary 2020, the AI con­trol sys­tem saved 27 per­cent of heat­ing ener­gy, com­pared to the neigh­bor­ing stu­dent room, whose heat­ing was oper­at­ed with a per­ma­nent­ly pro­grammed (rule-based) con­trol system.

“The beau­ty of our self-learn­ing AI con­trol sys­tem is that it can be used not only in the NEST research build­ing, but also in any oth­er build­ing,” says Bratislav Sve­tozare­vic. “You don’t need an engi­neer to pro­gram the con­trol, or any­one who ana­lyzes the house before­hand and cal­cu­lates a tai­lor-made solution.”

Pleas­ant warmth in a eco­nom­i­cal way

In a next step, Sve­tozare­vic and his col­leagues now want to deter­mine how the sys­tem can be expand­ed from a room to larg­er build­ings. “In our first exper­i­ment, we want­ed to depict a typ­i­cal house­hold of the future,” says the Empa researcher. For the sake of sim­plic­i­ty, the team has lim­it­ed itself to heat­ing and charg­ing. The work, how­ev­er, lays the foun­da­tion for much more. Sve­tozare­vic is cer­tain: “Our AI con­trol sys­tem can still cope even when a pho­to­volta­ic sys­tem sup­plies elec­tric­i­ty, a heat pump and a local hot water stor­age sys­tem have to be oper­at­ed – and the com­fort require­ments of the res­i­dents change again and again.”

How­ev­er, a new gen­er­a­tion of elec­tric cars is need­ed to use the AI sys­tem for opti­mal ener­gy sup­ply in the future. The cur­rent Euro­pean and US mod­els with the CCS quick-charg­ing con­nec­tion can only refu­el, but not deliv­er one. Japan­ese cars with Chade­mo plugs, on the oth­er hand, are designed for so-called bidi­rec­tion­al charg­ing. Kore­an group Hyundai announced in Decem­ber that it would also equip its new E‑GMP elec­tric car plat­form for bidi­rec­tion­al charg­ing. This would allow elec­tric cars to help save ener­gy in the long term and at the same time sta­bilise the elec­tric­i­ty grid.

More infor­ma­tion on the sub­ject can be found at: