A 6% Smarter Solar Forecast Slashed Grid Injections by 82%

An 82% reduction in grid injections. That's the number to sit with. In a grid-connected solar microgrid on the rooftop of a Norwegian university, a smarter way of predicting sunshine — combined with a controller that treats the battery, the grid, and the weather as one integrated puzzle — cut the sudden floods of midday solar power heading back into the electricity network by more than four-fifths. Not by building anything new, not by adding panels or batteries, but simply by deciding better when the power should be stored, sold, or consumed.

This is the quiet frontier of the clean energy transition. The easy part of solar power is generating it; the hard part is that the sun doesn't care when you need electricity. It peaks at solar noon, often exactly when demand doesn't. On a cloudy day, output collapses. Left unmanaged, a neighborhood full of solar roofs becomes a source of congestion and voltage swings that grid operators must spend money to fix. A team at the Norwegian University of Science and Technology (NTNU), led by Berhane Darsene Dimd with Steve Völler and Ole-Morten Midtgård, asked whether smarter software could smooth that chaos out — and the answer, published on arXiv, is a resounding yes.

The study pairs a long short-term memory (LSTM) neural network — a type of machine-learning model built to remember patterns across time — with a multi-objective control policy that schedules when a battery charges and when power flows to the grid. The two working together raised the fraction of solar power consumed locally from 78.1% to 84.5%, slashed grid injections by 82%, and shaved the forecast error by 6% relative to a naive baseline. The catch: that improved performance comes at the cost of more battery cycling, which accelerates battery aging. It's a trade-off the researchers deliberately built into the objective function, and it points toward what's likely the central engineering question of the distributed solar era — how to balance local resilience, grid health, economics, and battery longevity against each other.

The Science

At the heart of the system is a grid-connected microgrid: a solar array, a battery energy storage system (BESS), a load that must be met at all times, and a connection to the main grid. The microgrid sits on the rooftop of NTNU's Department of Electric Energy in Trondheim, Norway, a 20 kWp (kilowatt-peak) plant. Every 15 minutes, the system must decide: does the battery charge, discharge, or idle? Does the microgrid buy from the grid, sell to it, or stand alone? These decisions are the "energy scheduling" (ES) problem.

The controller that makes these choices is built on model predictive control (MPC), a feedback method widely used in process control. The idea is elegant: rather than reacting to what just happened, the controller uses a model of the system to predict the future, solves an optimization problem over a horizon (here, 24 hours ahead in 15-minute steps), implements just the first decision, then shifts forward and repeats. It's the difference between driving while looking at the road ten meters ahead versus planning your route from a map.

The optimization itself is formulated as a mixed integer linear program (MILP) — a type of mathematical problem that combines continuous decisions (how much power flows) with binary on/off decisions (is the battery charging or discharging this step?) — solved in Python with the Gurobi solver. Because the battery's state-of-charge equation is nonlinear (it behaves differently charging than discharging), the authors carefully reformulate it into a set of linear constraints using auxiliary variables, a standard but non-trivial piece of mathematical surgery that keeps the problem solvable in real time.

The objective function is where the multi-objective philosophy lives. It's a weighted sum of four competing goals: minimize the cost of buying grid electricity (using real Nord Pool spot prices as inputs); penalize every unit of energy imported from the grid, to discourage reliance on it; penalize the frequency and intensity of battery use, to slow degradation; and shape behavior to avoid dumping large surges of PV power into the grid during solar noon. The last goal is subtle and important: it doesn't just prevent grid injection outright, but actively discourages the timing pattern that causes midday congestion problems, by making battery cycling cheap enough to absorb the peak.

The forecasting component is an LSTM network. Where a conventional recurrent neural network tends to "forget" distant inputs — a problem called the vanishing gradient, where weight updates shrink to insignificance — LSTM units contain gates that regulate the flow of information into, out of, and within a memory cell, letting the model hold onto long-term patterns. The architecture was tuned iteratively on a validation set to settle on two hidden layers of ten units each, a batch size of 24, 150 epochs, and a learning rate of 0.001. Crucially, the model uses a lookback window of 96 steps (24 hours) to predict the next 96 steps — meaning it has a full day of history and forecasts a full day ahead.

The data comes from a full year of 15-minute PV measurements from 2020, with incomplete and winter months (which have near-zero output at high latitude) excluded, leaving 153 days for training and validation and 61 for testing.

What They Found

To meaningfully evaluate the forecasting's value, the team ran three scenarios against the same MPC controller: a perfect forecast (an idealized upper bound where the controller sees the future exactly), a persistence model (a naive baseline that assumes the forecast at time $t+1$ equals the observed value at time $t$), and the LSTM-based forecast.

The LSTM beat persistence on forecasting accuracy, cutting root mean squared error (RMSE) by 6% (Dimd et al., 2026).

Forecast error drops 6% with LSTM

Label	Value
RMSE (relative)	100
RMSE with LSTM	94

The LSTM forecast cut RMSE by 6% compared with the persistence model.

shows this comparison. RMSE is a measure of average prediction error, weighted so that large misses are penalized more heavily — it's the standard yardstick for forecast models. A 6% improvement may sound modest, but because the LSTM is being embedded in a downstream optimization, even small gains in accuracy can compound into large gains in operation quality.

That amplification is exactly what the researchers found. Across the three scenarios, the PV self-consumption ratio — the fraction of solar energy generated that is consumed locally rather than exported — rose from 78.1% with the persistence model to 84.5% with the LSTM forecast (Dimd et al., 2026).

Local solar consumption rises with better forecasting

Label	Value
Persistence model	78.1
LSTM forecast	84.5

Self-consumption rose from 78.1% with the persistence baseline to 84.5% with the LSTM-based forecast.

plots this. That's a 6.4 percentage-point improvement in local solar use, driven entirely by better prediction feeding smarter scheduling.

The biggest headline, though, is grid injections: the LSTM-driven system reduced power sent back into the grid by 82% compared to the persistence-based operation. This is the operational flip side of the same coin. When the controller knows more accurately how much sun is coming, it can time battery charging to absorb solar peaks rather than letting them spill onto the grid, and time its grid purchases for low-price windows instead of importing reactively. The result is a microgrid that behaves far more politely toward the electricity network — exactly the behavior grids need as solar penetration grows.

The paper also makes its methodological point quietly but clearly. By comparing against the perfect-forecast scenario, it establishes what the ceiling of performance looks like, and therefore how much headroom remains for better forecasting. That framing — an idealized upper bound, a naive baseline, and a realistic neural model in between — gives an honest sense of where the value in forecasting actually sits.

Why This Changes Things

There's a temptation to read this as a niche optimization exercise, a matter of engineering elegance rather than substance. That would be a mistake, because the substance here is precisely the problem that's about to hit every grid on Earth.

The story of solar power's rise has, so far, been mostly a story of manufacturing and cost: panels got cheap, installations boomed. But the next phase is about integration. When renewable penetration grows, the challenge stops being generating clean electricity and becomes managing its variability. The midday surge — when all the solar roofs on a street peak simultaneously while demand is low — is not a hypothetical. It's a measurable, growing phenomenon that grid operators across Europe, Australia, and parts of the US increasingly describe as a strain on voltage regulation and frequency stability. The conventional remedy has been to build more infrastructure, or to curtail — to simply switch off some solar generation when there's too much of it, which is economically and morally wasteful.

This study demonstrates an alternative that costs almost nothing by comparison: better prediction plus smarter control. An 82% reduction in grid injections on a single microgrid is a proof of concept that software can absorb much of the integration burden. Scaling that logic across many distributed systems, each deciding with better information when to charge, sell, or hold, would transform how much strain distributed solar puts on the network. The researchers explicitly frame their savings-control objective — discouraging the timing pattern that causes congestion rather than just minimizing injection volume — as a way to spare grid operators the expense of curtailment infrastructure. That's not an academic abstraction; it's a direct economic argument for why accurate forecasting pays.

The self-consumption finding matters for a more personal economic reason. In many electricity markets, the price paid for exporting solar power to the grid (often a feed-in tariff or wholesale spot price) is far lower than the retail price paid for importing it. So every kilowatt-hour consumed locally instead of exported is worth more to a prosumer. Raising self-consumption from 78.1% to 84.5% is a concrete improvement in the economics of owning solar — the kind of gain that, across millions of installations, reshapes the payback calculation for rooftop PV.

But the paper is admirably honest about the cost of these gains. The battery cycling penalty in the objective function was deliberately tuned so that the controller would use the battery more aggressively to capture solar peaks and arbitrage spot prices — and that extra throughput accelerates battery degradation. The researchers note this tension directly: improved operational performance and higher battery throughput travel together. This is the fundamental trade at the center of battery-backed solar, and it has real economic weight, since batteries are often the single most expensive component of a microgrid and their lifetime is typically measured in cycles, not years. The design choice to expose this trade explicitly — rather than hide it — is a strength of the work. It frames the question not as "is battery storage good?" but as "how aggressively should any given system use its battery, given its price, its chemistry, and the local market conditions?"

It's also worth noting where the model's mathematical elegance does real work. The reformulation of the battery's nonlinear charging behavior into a linear program isn't just an academic flourish — it's what makes real-time control possible. A system that can't be solved in the minutes available isn't a controller, it's a simulation. By keeping the problem a MILP, the authors ensure the decisions can actually be computed and implemented on a rolling 15-minute basis. This is the difference between a paper that describes a good idea and a system that could plausibly run a real building.

What's Next

The authors are explicit about the limitations and the road ahead, and both are instructive. First, the forecast here is deterministic — it produces a single predicted value for each future time step, with no measure of its own uncertainty. Cloud cover is famously hard to pin down hours in advance, and a forecast that is confident-but-wrong can mislead the controller just as badly as a naively imprecise one. Their stated next step — probabilistic forecasting, which produces a distribution of possible futures rather than a single line — matters because it lets the MPC weigh risk. A controller that knows a cloudy afternoon is possible but not certain can hold more battery reserve than one that assumes the deterministic prediction is right.

Second, the current framework forecasts PV but assumes the load is known. Real microgrids don't have that luxury; consumption patterns swing with occupancy, weather, and behavior. The authors flag load prediction as an explicit future direction, and it's a fair one — a system that mis-predicts both supply and demand has twice the error surface to navigate.

Third, and perhaps most interesting for the grid-level story, they plan to develop "smart control strategies that allow grid-support functionalities from the PV side." That phrase signals a shift in mindset that's quietly happening across the industry: rather than treating inverters and storage as passive loads on the grid that must be managed for, treating them as active participants that can provide services to the grid — voltage support, frequency response, reactive power. A microgrid that can not only avoid congesting the grid but actively help regulate it is the logical endpoint of this trajectory.

There are also caveats the paper itself acknowledges in spirit. The scenario-based comparison is built on retrospective data, and real-world performance under live conditions — with forecasting errors accumulating, spot prices moving unexpectedly, and hardware constraints biting — will be messier than the simulation. The two-month simulation horizon is short for drawing conclusions about battery aging or seasonal performance patterns. And the specific magnitudes (6% RMSE improvement, 82% injection reduction) are for this plant, this location, this battery chemistry, and these penalty weights; they won't transfer numbers-identically to other settings.

But the architecture transfers, and that's the point. The study demonstrates, rigorously, that forecasting accuracy and control policy are not separable technical details but two halves of one decision system — and that improving the first pays dividends through the second far out of proportion to the raw forecast gain. The 6% better forecast produced an 82% reduction in grid injections and a meaningful bump in self-consumption. That's leverage, and leverage is what the energy transition needs.

The deeper significance is a reframing of what "smart" means in the grid of the future. For decades, the dominant model was centralized: big power plants, one-way power flow, control from the top. Distributed solar and battery storage are forcing a shift to a world where millions of small, variable actors must coordinate. The naïve hope is that more hardware solves the problem — more batteries, more transmission, more grid infrastructure. This paper suggests that much of the solution is information. When each actor sees the future more accurately and decides more skillfully, the whole system runs better without anyone building a single new piece of equipment.

The sun will keep being unpredictable. The good news is that predicting unpredictability is getting cheaper, faster, and better all the time.