AI predicts short-term electricity demand in solar-powered homes

The rooftops of the world are quietly turning into power stations, and the shift is breaking one of the oldest assumptions of the electricity business: that a utility can predict, with reasonable confidence, how much power a home will draw. Once a house installs solar panels, it becomes a consumer at night, a miniature power plant at noon, and something far less predictable whenever a cloud drifts overhead. A new study from researchers in Cyprus reports that a specialized class of artificial neural networks can tame this uncertainty, forecasting the net electricity flows of solar-equipped homes with errors that fall to just a few percent — far better than the simple persistence rules that grid operators have long leaned on as a safety net. Writing in the journal Energy Reports, Georgios Tziolis and his colleagues describe how a Bayesian neural network, trained on a single year of measurements from 68 real households, consistently beat the baseline approach for individual homes and entire neighborhoods alike.

The timing could hardly be more pointed. According to the International Energy Agency, global renewable electricity generation is forecast to exceed 17,000 terawatt-hours by 2030 — an increase of almost 90 percent over 2023 — with renewable sources supplying 46 percent of the world’s electricity that year. Solar photovoltaics, the technology propelling much of that growth, is on track to become the largest single renewable power source by 2029. Behind the meter, the same revolution is gathering pace: falling equipment and installation costs, down more than 80 percent over the past decade, have pushed the price of rooftop solar to roughly one dollar per watt. Around 25 million households worldwide relied on rooftop solar power in 2022, and the IEA expects that figure to climb past 100 million by 2030. In the European Union the pressure is regulatory as well as economic, because under the EU Solar Energy Strategy, rooftop solar will be mandatory for all new residential buildings from 2029.

All of this turns net load — the difference between what a building consumes and what its panels generate — into one of the most consequential and slippery quantities in modern power systems. A conventional load curve is governed by human routine: alarm clocks, kettles, washing machines, evening lights. A net load curve superimposes the whims of the sky. When a photovoltaic array produces more than the household needs, the net load turns negative and power flows back into the grid; when clouds throttle the panels during an evening demand spike, the shortfall must be covered from elsewhere, instantly. In the Cypriot dataset, one of the homes was on average a net exporter of electricity across the whole year, with a mean net load of minus 0.07 kilowatts — a profile that would have been almost unthinkable for a residential customer a generation ago. Forecasting this quantity hours ahead, a task known as short-term net load forecasting, is what allows network operators to balance supply and demand, schedule resources, and keep the lights on without expensive last-minute interventions.

The new work expands on an earlier proof of concept by the same team, which tested a Bayesian neural network on data from just six Cypriot homes. For the expanded study, the researchers assembled a full year of measurements from 68 households with photovoltaic systems and widely varying consumption habits. Each home’s net load was recorded every 30 minutes and then averaged to hourly values, producing two parallel datasets at 30-minute and 60-minute resolutions — intervals chosen to match international photovoltaic monitoring standards and the operational tempo of home energy management systems and utility scheduling. Every forecast drew on six inputs: the net load recorded at the same time of day one week earlier, air temperature, solar irradiance, the time of day, the day of the week, and the month of the year. Earlier analyses by the group, using Pearson correlation and mutual information, had singled out these variables as the most informative predictors. The raw records were first scrubbed of duplicates, gaps, and sensor faults using established data-quality procedures for photovoltaic systems.

Before any forecasting began, the team faced a deceptively simple question: which homes deserve close scrutiny? To answer it, they turned the 68 households into a map. A self-organizing map — an unsupervised neural technique that compresses many dimensions of data onto a two-dimensional grid, trained here with a learning rate of 0.01 over 10,000 epochs — was combined with a mean shift clustering algorithm that scans for density peaks using kernel density estimation and gradient analysis. The procedure revealed four distinct clusters, defined by three telling variables: the ratio of total electricity consumption to total photovoltaic production, the average net load, and the maximum load. Twenty homes with low values across all three variables formed cluster A; twenty more with medium ratios and medium-to-high peaks made up cluster B; sixteen homes with medium ratios but modest peaks composed cluster C; and twelve power-hungry households with high consumption-to-PV ratios filled cluster D. From these groups the researchers selected nine representative homes, at least two per cluster, spanning consumption-to-PV ratios from 0.89 to 2.94.

The forecasting engine at the heart of the study is a Bayesian neural network: a three-layer network whose predictions are framed probabilistically, allowing the model to handle uncertainty in the data rather than pretending it does not exist. Six input nodes feed eleven hidden neurons — a size governed by a rule of thumb that keeps the hidden layer smaller than twice the input layer — and a single output node delivers the net load forecast. The Bayesian formulation was a deliberate choice for this problem, because the one-year dataset is small by deep learning standards and laced with volatility, and previous work by the team had shown the approach outperforming conventional artificial neural networks and support vector regression on the same task. Bayesian networks also train quickly and resist overfitting on short datasets. The model learned from a random 70 percent of the yearly data and was tested on the remaining 30 percent, racing against a deliberately humble opponent: the naïve persistence model, which simply assumes the next interval will resemble the same interval one week earlier.

Across the nine individual homes, the Bayesian network posted daily mean errors, normalized by each home’s maximum measured net load power, of between 5.44 and 8.00 percent on the 30-minute data and between 5.31 and 9.06 percent on the hourly data. The persistence model managed only 7.43 to 10.72 percent and 7.30 to 12.89 percent over the same tests. Expressed as a skill score — the standard percentage improvement over the naïve baseline — the neural network won every contest, with gains of 20.37 to 36.71 percent at half-hourly resolution and 22.45 to 37.84 percent hourly. The pattern within the sample was instructive. The toughest customer was household 1, the home with the lowest consumption-to-PV ratio and the lowest overall demand, the very net exporter whose small, solar-dominated profile swings wildly with the weather. The strongest result came from household 6, with the highest ratio of 2.94 and the highest average net load, where steady demand hands the model a firmer signal to learn from.

The most striking results emerged when the researchers stopped examining houses one by one and summed all 68 into a single aggregated profile — roughly the view a distribution network operator actually holds. At 30-minute resolution, the Bayesian model’s root mean square error for the aggregate was 10.72 kilowatts against 19.99 kilowatts for persistence, a reduction of 9.27 kilowatts, nearly half. Its normalized error was 7.20 percent versus 11.96 percent, a skill score of 46.37 percent. On hourly data the gap widened further: 9.97 against 19.62 kilowatts, 6.77 against 11.52 percent, and a skill score of 49.18 percent, the highest recorded anywhere in the study. Aggregation, the authors explain, smooths away the random flicker of individual households — a dishwasher here, a kettle there — leaving the underlying rhythm of collective consumption and solar production, which the neural network captures far better than a copy-last-week rule. The persistence approach fared worst on the aggregated hourly profile, where its daily error spiked to 42.87 percent on one turbulent day; the neural network also slipped above 10 percent on a few extreme-weather days, but never remotely approached such collapses.

Because solar variability is ultimately a weather story, the team stress-tested the model across three days with sharply different irradiance profiles. On the dimmest day, with a mean daily irradiance of 172 watts per square meter, the Bayesian network’s normalized error was 5.49 percent against 6.52 percent for persistence. On a middling day averaging 225 watts per square meter, the baseline sagged to a 10.44 percent error while the neural network held firm at 5.35 percent. And on the brightest day, averaging 276 watts per square meter with peaks near 966, the Bayesian model delivered its best figure of the entire evaluation: 3.95 percent, against 5.84 percent for the baseline. The pattern is counterintuitive but logical. The neural network’s forecasts of aggregate net load actually sharpened as the sun strengthened, while the naïve model, blind to everything except last week’s numbers, drifted badly during the bright daytime hours when photovoltaic output swings were at their most violent.

Speed matters as much as accuracy for anything that must run inside a live forecasting platform, and here the results are equally favorable: the Bayesian network completed both training and testing in under a minute, whether for a single home or the full 68-household aggregate. The authors argue that the model can be slotted directly into forecasting tools and energy management platforms to support operator decisions across the full spectrum of consumption-to-PV ratios and load profiles. The research was carried out within the DSM4islands project under the CETPartnership, co-funded by the European Commission and national programs in Germany, Cyprus, and Italy. Next on the agenda, the team says, are validations on households in different climates, benchmarks against heavyweight deep learning architectures such as long short-term memory networks, gated recurrent units, and transformer models, and a careful quantification of the trade-off between forecasting accuracy and computational cost. With regulators mandating rooftop solar on every new European home and more than 100 million solar households expected worldwide by 2030, the unglamorous chore of predicting a neighborhood’s net electricity flow is quietly becoming one of the modern grid’s most valuable skills.