The Challenge
Fysiko Aerio Ellados (formerly Attiki Gas Supply Company — EPAA) is Greece’s largest city-gas distributor, supplying natural gas across the Attica basin to 330,000 households, 200 large commercial and industrial clients, 6,500 businesses, and 1,200 schools and public buildings.
As a shipper in the National Natural Gas System (NNGS) — the high-pressure backbone transmission network operated by DESFA — Fysiko Aerio Ellados carries a precise regulatory obligation: every day, it must declare its capacity booking for the following day. This declaration is a commitment — a forecast of how much gas its customers will collectively consume.
The NNGS rules are unforgiving on accuracy. Positive deviation (consuming more than declared) and negative deviation (consuming less) both trigger financial imbalance charges. Over the course of a year, small systematic errors in daily forecasts translate directly into material penalties. The accuracy of the daily demand estimate is not an operational detail — it is a core driver of the company’s cost position.
The fundamental difficulty is that natural gas demand is deeply weather-sensitive. Residential heating loads dominate consumption in Attica, meaning that a single cold front sweeping through Athens can swing daily gas demand by tens of millions of cubic metres. Simple historical averages or degree-day models are too coarse for the precision that daily NNGS declarations require. What the forecasting system needed was high-quality, high-resolution weather data — not generic regional forecasts, but data tailored to the spatial distribution of customers across the service territory.
Why Weather Data Quality Is Everything
Machine learning models for energy demand forecasting are only as good as the data they are trained on and the real-time inputs they receive. Two factors matter most:
Spatial coverage. The Attica basin spans a large and topographically varied area — from coastal zones to inland suburbs and mountain-facing districts. Temperature and wind patterns differ meaningfully across these zones, and gas consumption tracks those local variations. A single representative station misses the geographic spread of demand.
Historical depth and consistency. Training a robust ML model requires years of clean, gap-free weather time series. Discontinuities, sensor failures, or sparse early records introduce noise that degrades model accuracy in exactly the edge-case conditions — severe cold spells, unusual wind patterns — where precise forecasting matters most.
Fysiko Aerio Ellados had the forecasting model. What it needed was a partner who could supply the right weather data infrastructure underneath it.
Our Approach
Ex Machina’s engagement began with a systematic analysis of the service territory: which measurement locations best capture the weather variability relevant to gas consumption across Attica? Using historical data and consumption patterns, we identified the optimal set of monitoring points — a network of locations whose combined readings provide the best proxy for the demand-driving meteorological conditions experienced by Fysiko Aerio Ellados’s customer base.
Historical data foundation. We assembled multi-year historical weather time series for each identified location — cleaned, quality-controlled, and delivered in the exact format required for ML model training. This gave the forecasting team the consistent, long-duration dataset needed to train models that generalise well across the full range of weather conditions the Attica basin experiences.
Operational data delivery. Once the model was trained and validated, Ex Machina transitioned to production delivery: a daily automated feed of processed, forecast-ready weather data for all relevant locations, arriving before Fysiko Aerio Ellados’s declaration window. Delivery is format-agnostic — API, web service, email, or custom integration — adapting to the operational workflows of the client rather than the other way around.
Ongoing model support. Gas consumption patterns evolve as the customer base grows and the building stock changes. Ex Machina supports experimentation with updated model inputs and alternative weather feature sets, enabling the forecasting team to continuously refine accuracy rather than locking in a static solution.
A Partnership Built on Responsiveness
The collaboration has been live since 2016. In that time, the demands on the weather data service have evolved — new customer zones, updated model architectures, changing delivery formats. Every adjustment has been handled quickly and transparently, without disruption to the daily production pipeline.
The value Ex Machina provides is not a static data feed. It is a combination of data quality, delivery reliability, processing flexibility, and the willingness to move fast when the client’s model or operational needs change. As the team at Fysiko Aerio Ellados has noted directly, it is this combination — not just the data itself — that makes the partnership work. Their own team captured it on record in a video testimonial.
Results
The link between forecast accuracy and financial performance at Fysiko Aerio Ellados is direct and quantifiable: every percentage-point improvement in the accuracy of the daily capacity booking reduces exposure to NNGS imbalance charges. Since 2016, the Ex Machina weather data service has been a consistent, reliable component of that forecasting pipeline.
Beyond the direct cost impact, the partnership has delivered something harder to quantify but equally important: confidence. When the most challenging weather conditions arrive — the cold fronts that drive peak demand, the unseasonal warmth that collapses it — the forecasting system has the data quality and coverage needed to respond accurately.
The model that Fysiko Aerio Ellados operates is theirs. The insights it produces are theirs. Ex Machina’s role is to ensure the weather data underpinning it is never the weak link.