If you’ve ever wondered why clean solar power isn’t yet flowing through every home and business in Europe and Canada, part of the answer lies in timing. We don’t know at what times the sun will shine , allowing solar panels to deliver. But we know that our electricity grid must always balance supply with demand. That’s where forecasting comes in.
A team at Edge Hill University in Ormskirk, United Kingdom [53.6°N, 2.9°W] has developed a new way to predict how much electricity solar panels will generate — not only with greater accuracy, but also in real time. Their tool, a hybrid artificial intelligence model called the BiLSTM-Informer, could improve how we use solar power across entire regions, helping to balance the grid and reduce dependence on fossil fuels.
This isn’t just a technical breakthrough for data scientists. It’s a practical step towards smarter, cleaner, and more reliable electricity — one that speaks directly to the challenges shared by Canada and Northern Europe alike.
Why Forecasting Solar Power Is So Important
Solar electricity is beautifully clean and increasingly cheap. But it’s also intermittent. Clouds, seasons, time of day — they all affect generation. If utility companies don’t know what’s coming, they must hold backup gas turbines or overbuild storage. The better our forecasts, the more efficiently we can use what we produce.
For years, engineers have used models that rely on historical weather data and machine learning to predict solar output. These models have steadily improved. But as the Edge Hill team points out, many of them still struggle to capture the full complexity of solar behaviour — especially over longer forecasting periods. They also tend to operate offline, rather than in the real-world conditions where grid operators need fast, accurate updates.
How Is This New Model Different?
The BiLSTM-Informer hybrid takes a more nuanced approach. It combines two powerful deep learning techniques:
- Bidirectional LSTM (Long Short-Term Memory) — which can remember and interpret patterns in time-series data both forwards and backwards.
- Informer — a more recent innovation that focuses attention on only the most relevant data points, reducing the noise and speeding up computation.
Together, they offer both accuracy and efficiency, even for long-range forecasts. But that’s not all.
The model also uses Fourier transformations and cyclic encoding — think of these as tools to recognise repeating seasonal and daily patterns — and autoregressive optimisation, which helps the system learn from previous outputs. The result is a model that learns how solar behaves in context, including all the quirks of British weather.
Even better, the team has deployed it online, with a web interface that runs in real time. You can try it for yourself: View the model here.
How Accurate Is The Model?
Compared with previous models (including popular tools like Lasso regression, K-nearest neighbours, and even classic deep learning like GRUs or LSTMs), the BiLSTM-Informer delivered:
- R² of 0.952 — a very strong indicator of predictive accuracy
- Mean Absolute Error of 1.22 kWh
- Real-world forecasting accuracy between 89% and 97.3%, depending on time horizon
Even at monthly scales, where many models begin to falter, this hybrid system maintained high performance. And because it’s been tested on five years of solar data from a real installation, it reflects the kinds of variability that real-world systems face — not just idealised lab conditions.
Benefits
The challenges this model addresses — fluctuating solar generation, integration with the grid, reliance on forecasting for dispatch decisions — are shared across Canada and much of Europe.
- In Canada, provinces like Alberta and British Columbia are scaling up solar while still depending on fossil fuel peaker plants. Better forecasting means less need for those.
- In Denmark, Germany, and the Netherlands, dense grid infrastructure already enables high renewable penetration, but the balancing act is constant. More accurate forecasts mean smoother energy markets and lower costs.
Moreover, because the model is designed to be web-deployable and lightweight, it could be integrated into grid management software without huge computational burdens. This is important for smaller utilities or community energy groups who may not have access to high-end modelling teams.
Smart Energy Needs Smart Tools
This isn’t a flashy gadget or a one-off research paper. It’s a practical demonstration of how far intelligent forecasting can go in helping us use renewable energy better. With reliable solar forecasts, grid operators can plan ahead, energy traders can reduce risk, and policymakers can build systems that genuinely support a low-carbon transition.
The work done at Edge Hill may not make many headlines. But for those of us looking at the nitty-gritty of how to live more sustainably — how to power homes, charge electric cars, or run a clean grid — this kind of precision, automation, and realism is exactly what’s needed.
Source
Real-Time Web Inferencing of a BiLSTM-Informer Hybrid Model with Autoregressive Features Optimization for Improved Photovoltaic Power Output Forecasting, SSRN Preprint, 2025-04-12
Boˆreal 