Beyond Rule-Based Energy Management
A first-generation energy management system runs on rules: charge the battery when solar exceeds load, discharge when load exceeds solar, stop charging at full. These rules work but they do not optimise — they react to current conditions without anticipating future ones. The practical consequence: a rule-based system does not know that cloud cover is forecast for tomorrow afternoon, that the grid tariff will be at its weekly high in four hours, or that a fleet of vehicles is scheduled to depart before 6 am and will need priority charging from 1–4 am when tariffs are lowest.
A machine learning EMS addresses this by modelling the energy system's future state and making dispatch decisions today based on a 24–72 hour optimisation horizon rather than the current moment. The improvement in outcomes depends heavily on the predictability of the site's load and generation patterns — sites with consistent daily patterns see the most benefit, while highly variable commercial sites see more modest gains.
What the ML Models Actually Do
An AI energy management system typically comprises three forecasting components and an optimisation layer:
- Solar generation forecast: Machine learning models combining historical irradiance data, weather API forecasts, and (where available) sky camera imagery can forecast generation to ±8–12% RMSE at a 15-minute resolution 24 hours ahead. This is accurate enough to base dispatch decisions on, though not precise enough to eliminate real-time correction.
- Load forecast: LSTM or gradient boosting models trained on site-specific load history can predict demand by 15-minute interval, capturing weekly patterns, seasonal effects, and known scheduled events. Sites with 6–12 months of operating history achieve meaningful forecast accuracy; new sites require a learning period.
- Tariff optimisation layer: Given generation and load forecasts, a constrained optimisation (typically linear programming) calculates the BESS charge/discharge schedule that minimises grid energy cost subject to battery constraints, sanctioned load limits, and vehicle departure requirements.
Published research on predictive versus rule-based energy management — from pilots in comparable grid and solar conditions — suggests 12–22% reductions in grid energy costs are achievable with well-implemented predictive dispatch, compared to rule-based systems managing identical hardware. The range reflects site variability; the average is meaningfully above zero, which is why the industry is moving in this direction.
What EVBooth's EMS Does Today — and Where It Is Going
EVBooth's EMS currently implements smart scheduling for EV chargers — coordinating session timing with tariff windows, BESS state, and solar generation using OCPP 2.0.1 composite schedules. The predictive forecasting layer is in active development, with solar generation forecasting using weather API integration at a prototype stage. The full AI-optimised dispatch capability — with load forecasting and closed-loop optimisation — is part of our product roadmap for 2026, timed to coincide with our Hybrid Loop site launches where we will have the multi-asset complexity (solar + BESS + chargers + fleet schedules) that makes predictive optimisation most valuable.
A Note on Realistic Expectations
AI energy management is genuinely valuable, but the industry has a tendency to overstate results. Vendor claims of 30–40% energy cost reductions should be scrutinised: what was the baseline, what hardware was it managing, and how were savings calculated? A well-designed ML EMS at a site with co-located solar, a BESS, and predictable load patterns can realistically deliver 15–20% improvement over rule-based management of the same hardware. That is a meaningful contribution to project economics — but it is not a substitute for right-sized hardware, correct tariff structure, and adequate site utilisation.