AI in Smart Home, Part 7: A Home That Plans - Energy Optimization with Forecasts
The seventh part of the AI in smart home mini-series. We build an energy planner: electricity prices, solar production forecasts, and the calendar as inputs, an LLM as a constrained planner, a decision log, and an honest profitability calculation.
Your home knows the weather forecast, tomorrow’s electricity prices, and your calendar. That is enough for AI to lay out a daily plan for the water heater, car charging, and heating - and to be able to explain every decision. In this part, we build an energy planner: from input data, through a decision log, to an honest calculation of when it pays off at all.
Three data sources the home already has or can easily get
Energy planning stands on three legs. The first is the price of electricity over time. With a dynamic tariff, the home may receive prices for the next day-ahead market blocks, but publication time and conversion depend on the supplier. In 2026 the settlement interval can be 15 minutes, so do not assume whole hours. With a regular time-of-use tariff, it is simpler - the cheaper windows are written into the contract.
The second leg is the production forecast, if you have photovoltaics. Integrations like Forecast.Solar convert weather forecasts and system parameters into expected production, with uncertainty that must be allowed for. Energy from the roof is not literally free: using it has an opportunity value determined by how exported surplus is settled.
The third leg is the calendar and household patterns. Leaving at 6:30 means the water has to be hot by 6:00 and the car charged in the evening, not in the early morning. Working from home means a higher temperature in the office during hours when the home would normally be saving.
What can really be shifted
Here comes the first dose of realism: most of your consumption cannot be shifted. The fridge, lighting, stove, TV - those happen when they happen. What can be shifted are appliances that store the effect of their work:
- the water heater - it stores hot water, so it can heat at 2:00 AM instead of 6:00 PM;
- the electric car - it stores energy and often provides a multi-hour charging window before the next trip;
- heating with thermal inertia - underfloor heating or a well-insulated house allows overheating the building by 1 degree during a cheap hour and waiting out the expensive one;
- the washing machine, dishwasher, dryer - a cycle can be delayed by a few hours with no loss of comfort at all.
The share of these categories depends on heating, the car, household size, and habits. Measure it before buying controllers, because it determines whether the project can pay off.
Architecture: the LLM plans, the automation executes
The pattern known from earlier parts of this series applies here too, in its purest form. The language model is a constrained planner: once a day, for example at 9:00 PM, it receives tomorrow’s price list, the production forecast from the roof, the car’s state of charge, the calendar, and a list of hard conditions. It returns a plan in structured form - a schedule with appliance start times plus a justification for each item.
Execution is handled by a classic, deterministic automation that knows the non-negotiable conditions and checks them independently of the plan: water at minimum 50 degrees by 6:00, the car at minimum 60% before the first departure in the calendar, the house temperature never below 19 degrees. If the model’s plan violates a boundary condition, the condition wins, not the plan. The model can be wrong, and guardrails can also be misconfigured - so they need tests, safe defaults, and a manual override.
Why an LLM at all, when off-the-shelf optimization algorithms exist? Honestly: for the water heater schedule alone, a plain “pick the 3 cheapest hours of the night” is enough. The language model brings two things: it combines sources a rigid algorithm does not know (the calendar, unusual situations, your notes about plans), and it explains decisions in human language. That second thing turned out to be more important for us than we expected.
The planner prompt: constraints matter more than the goal
From Grzegorz’s experience with agents at work comes one principle that transfers to the home one-to-one: a model optimizing without explicit constraints will optimize your life in a way you do not want. A planner told to “minimize energy cost” will happily leave you with cold water at 6:00 AM, because not heating was cheaper.
That is why the planner prompt consists mostly of constraints, not of the goal. The temperatures below are only comfort examples, not a complete hot-water hygiene program - configure Legionella protection for the installation and the manufacturer’s guidance:
- water at minimum 50 degrees between 5:30 and 7:00 and between 20:00 and 22:00;
- the car charged at minimum to the level from the calendar one hour before the first departure;
- room temperature never below 19 degrees, in the kid’s room never below 20;
- the washing machine and dishwasher do not start between 22:00 and 7:00 (noise), the dryer only when someone is home;
- at most 3 high-power appliances at the same time (grid connection limit);
- if the input data is incomplete, plan the safe variant and note it in the justification.
Only after the list of constraints comes the goal: within the remaining freedom, minimize cost and prefer your own production from the roof over buying from the grid. That order sounds pedantic, but it is what decides whether the plan is helpful or a nuisance.
The decision log, or trusting the machine
A home that switches appliances on and off by itself at night breeds distrust - right up to the moment it starts explaining itself. Every item of the plan goes into our decision log: a simple log on the dashboard, one sentence per decision.
21:04 PLAN for Tuesday 21.07:
- Water heater: 02:00-04:00 (price 0.34 zl/kWh, weak PV forecast: 2.1 kWh,
calendar: departure 6:30, water ready before 6:00)
- Car: 01:00-05:30 to 80% (cheapest window of the night, 90 km tomorrow per calendar)
- Dishwasher: start 13:00 (PV forecast peak, 3.4 kW from the roof)
- Heating: +1 deg C 05:00-07:00, setback 18:00-20:00 (price peak 1.12 zl/kWh)The log serves three functions. It builds household trust, because “why did the washing machine start at 1 PM?” has an answer on the dashboard. It makes debugging easier, because a bad plan is visible together with the data it was built on. And it disciplines the prompt - if the model cannot briefly justify a decision, that is usually a sign the decision is random.
Deviations should also be written into the log: every manual override of the plan by a household member becomes an entry “the plan called for X, the human chose Y”. A “heat now” button on the dashboard must exist and work without argument - a home where you cannot take a hot shower outside the schedule is a badly designed home, regardless of the savings. And if the deviations keep repeating, it is not the household that is the problem, but the constraints in the prompt that need fixing.
Let’s do the math: the synthetic DemoHouse
Take a fully synthetic DemoHouse: a 2+2 family, an electric car, an electric water heater, and a contract for which we assume an average 0.35 zl/kWh in selected cheap blocks and 1.10 zl/kWh in the expensive 17:00-21:00 window. These are assumptions for showing the calculation, not a current price list or every home’s typical profile. For simplicity, the table compares the variable energy purchase component; a full assessment must include the contract, distribution, tax, and fees.
| Appliance | Energy per year | Without a plan (avg. price) | With a plan (avg. price) | Savings |
|---|---|---|---|---|
| Car (15,000 km, 18 kWh/100 km) | 2700 kWh | 0.85 zl/kWh | 0.38 zl/kWh | approx. 1270 zl |
| Water heater (6 kWh/day) | 2190 kWh | 0.80 zl/kWh | 0.40 zl/kWh | approx. 880 zl |
| Washing machine and dishwasher (450 cycles at 1.1 kWh) | 495 kWh | 0.75 zl/kWh | 0.45 zl/kWh | approx. 150 zl |
If DemoHouse also has photovoltaics, a second lever appears: self-consumption of your own production. Every kilowatt-hour of dishwasher use shifted from evening to noon may replace a grid purchase, but you also give up the settlement value of that exported surplus. Calculate the increase in self-consumption from the production and load profiles, because the 5 kWp system size alone cannot support a promised percentage-point or monetary gain.
The table totals about 2300 zl per year, but only under its assumptions and before the complete bill. On the cost side, include controls rated for the water heater and charger, installation by a properly qualified person, any contactors and protection, the server, its electricity, model queries, and configuration maintenance. A local model has no per-query fee, but its hardware and electricity still cost money. Payback comes from the full comparison, not the relay price alone.
The limits of profitability
Now the second dose of realism. Cross the car out of the table - about 1000 zl per year remains. Cross out the dynamic tariff and keep a regular flat-rate tariff with a single price - zero remains, because there is nothing to optimize on price (only self-consumption of PV production remains, if you have it). A home without an electric car, without photovoltaics, and without a time-of-use or dynamic tariff does not need an energy planner - at most it needs part six of this series, that is, detecting that something is consuming too much.
A simple test before you start: calculate how many kilowatt-hours per year you can realistically shift and simulate two complete bills under your contract. Also test cases where the plan runs on, for example, 60%, 80%, and 100% of days instead of subtracting an arbitrary fixed percentage. If the result does not cover cost plus a margin for uncertainty, let it go or limit yourself to one fixed-schedule water-heater automation - no forecasts, no model, no log.
Summary
Energy planning is measurable because its effect can be checked on the bill. There are three ingredients - prices over time, the production forecast, and the calendar - and the architecture pattern is unchanged: the model proposes and justifies, while deterministic automation executes within tested conditions. The decision log makes the automation easier to understand. Calculate your shifting potential before you start - an electric car and variable pricing may increase it, but do not guarantee savings. If the numbers come out positive, start with one controlled load and verify the result on the bill before expanding the system.