Artificial Intelligence & Machine Learning
We build and deploy the models that turn load curves, price signals, and asset behavior into decisions traders and operators can act on — and we put the pipeline around them so they keep working after we leave, not just in the pilot.
Book a Consultation →Typical engagement deliverables
- Forecasting models (load, generation, price)
- Anomaly detection for consumption and trade surveillance
- Optimization models for dispatch and supply
- MLOps pipeline for retraining, monitoring, and drift detection
Technology & partner ecosystem
The method matters more than the model name.
The hard part of an energy ML model is rarely the algorithm — it is the features, the thresholds, and knowing when the model is wrong. These are the problems we have built and run most, and how we approach them.
Demand & load anomaly detection
We model expected demand from weather-driven features — temperature, humidity, seasonality — on interval meter data, then flag deviations statistically. We ensemble several models (regression, gradient boosting, random forests, Gaussian mixtures) rather than trust one, and tune residual thresholds to the false-positive cost the operator can actually live with.
Load & price forecasting
Short-horizon forecasts on interval data, feature-engineered for weather, seasonality, and calendar effects — then carried straight through to projected revenue and margin, so a forecast lands as a business number a desk can act on, not just a curve.
Customer & margin segmentation
Unsupervised clustering to group accounts by profitability and load behavior, so pricing, retention, and collections effort go where the margin actually is — not spread evenly across a book that is anything but even.
Consumption & trade surveillance
The same statistical backbone pointed at revenue leakage and compliance — surfacing the meter reads, usage, and trades that should not be there, scored and ranked so investigators start with the cases that matter.
Interactive demos built with AI & ML.
Live Power BI dashboards on synthetic data — click any to load it in place.
DER Penetration & Prediction
Impact of EV, energy-efficiency, and economic trends on consumption — synthetic data.
Wind Asset Performance & Weather
Correlate wind-asset output with detailed weather to find most- and least-profitable scenarios — synthetic data.
Solar Asset Performance & Weather
Correlate solar-asset output with detailed weather to find most- and least-profitable scenarios — synthetic data.
Heat Rate Options Analytics
Spark-spread and heat-rate option valuation for generation assets — synthetic data.
Supplier Performance Analytics
Benchmark and score fuel suppliers on cost, quality, and reliability — synthetic data.
Consumption Anomaly Detection
Spot anomalies across consumption levels and surface correlations with wider demographic trends — synthetic utility data.
Predictive Asset Maintenance
Assess asset performance, predict likely failures, and model upstream/downstream economic impact — synthetic utility data.
Where artificial intelligence & machine learning shows up in energy.
A power utility cut its load-forecast error from 18–25% down to 2.75%.
An 18–25% swing between load forecast and actuals meant costly real-time market exposure. The strategic target was under 4%. We delivered forecasting that brought average variance to 2.75%.
A battery operator can see the nodal spread before it's obvious — and dispatch against it.
We built NodalAlpha — a battery-storage arbitrage engine that forecasts nodal basis spreads, ranks every ERCOT and CAISO node by forward opportunity, and schedules charge and discharge against a sized unit. Backtested on a full year of real ERCOT settlement prices, energy arbitrage only.
A commodity trading firm made compliance proactive — catching trading aberrations before they became violations.
Trading commodities across global markets meant living under many overlapping regulations. We built transaction surveillance that flags untypical trading activity for investigation before it becomes an enforcement problem.
A renewables portfolio replaced fragile spreadsheet tax-equity modeling with a rules-based engine analysts could trust.
Analysts were spending their time manipulating spreadsheets to model complex wind tax-equity deals instead of analyzing them. We built a rules-based engine that captures the deal terms and does the modeling — so the team could get back to decisions.
A retail energy provider learned which customer and product mixes were actually profitable.
In deregulated markets, a complex mix of products and contracts made it hard to tell which customers and deals paid off over time. We built the analytics to evaluate profitability by product, customer, and deal — and the churn behind it.
A power marketer automated the gross-margin and variance analysis it used to assemble by hand.
Forecast-variance and realized/unrealized P&L analysis was manual and slow. We built the data foundation and reporting to automate gross-margin analysis, cash-flow forecasting, and budget-to-actual variance analysis.
How a artificial intelligence & machine learning engagement runs.
- 01
Assess
We start by understanding the current state — the data, the systems, and the real problem underneath the stated one.
- 02
Roadmap
We prioritize the work that will actually move the business, sequence it, and build the case for funding it.
- 03
Build
We deliver the pipelines, models, and reporting — working alongside your team, not in a silo.
- 04
Run
We hand over a solution your team can operate and evolve, with the enablement to keep it running after we leave.
Perspective on artificial intelligence & machine learning.
A backtest with a perfect record saw the future
A trading signal earns trust by showing where it's wrong — its losing days, its confidence band, its drawdown. A 100% hit rate isn't reassuring; it's evidence the model was allowed to peek at the answer.
Jul 25, 2026 · Read →The revenue you already earned: finding leakage in meter-to-cash
The cheapest revenue a utility can find is the revenue it already earned but never collected. Leakage hides between the meter and the bill — and the win is less about a smarter model than about attribution you can act on.
Jul 19, 2026 · Read →Churn scores are cheap. Knowing where to spend retention is not.
A churn probability on every customer is easy to produce and easy to waste. The harder, more valuable question is which of those customers are worth keeping — and that means tying churn to lifetime value, not just risk.
Jul 16, 2026 · Read →