Equipment failures that weren't predicted, grid imbalances found after the fact, and field technicians dispatched reactively: these are the operational costs that AI reduces in energy and utilities. The sensor and meter data to prevent them already exists in most operations.
We build AI systems for utilities, energy companies, and oil and gas operators: predictive maintenance for generation and distribution assets, demand forecasting for grid management, anomaly detection for pipeline and grid infrastructure, AI-powered field service routing, energy consumption optimisation for buildings, renewable energy output forecasting, and AI-driven customer billing anomaly detection. Every system is scoped against your operational data and a specific asset or cost target.
Predictive maintenance models that surface asset failure risk weeks before a trip or outage occurs
Demand forecasts at the feeder and substation level that reduce both reserve costs and grid imbalances
Anomaly detection on SCADA and pipeline sensor data that flags deviations before they become incidents
Renewable output forecasts that improve dispatch decisions and reduce curtailment costs
RaftLabs builds AI systems for utilities, energy companies, and oil and gas operators including predictive maintenance models for generation and distribution assets, demand forecasting for grid management, anomaly detection on SCADA and pipeline sensor data, AI-powered field service routing, energy consumption optimisation for commercial buildings, renewable energy output forecasting, and customer billing anomaly detection. Predictive maintenance models typically surface asset failure risk 2--4 weeks before a trip or outage. Each engagement is scoped at a fixed price after a discovery phase that maps your sensor data and maintenance records to the specific AI capability.
Trusted by
Operations that see asset risk before it becomes an outage
Energy AI is most valuable when it moves your operations team from reacting to failures to anticipating them. The sensor data to predict most equipment failures already exists in your SCADA and monitoring systems. The question is whether a model is turning that data into actionable maintenance signals.
Capabilities
What we build
Predictive maintenance models
Time-series models trained on your historical sensor data -- vibration, temperature, oil analysis, current draw -- combined with your maintenance records and failure history. Scores each asset by current failure probability and surfaces the contributing sensor signals. Gives maintenance teams a prioritised work list based on actual failure risk rather than calendar-based schedules. Applicable to transformers, switchgear, rotating machines, pumps, and pipeline equipment. Prediction horizon is typically 2-4 weeks before failure onset.
Grid demand forecasting
ML models trained on historical interval load data, weather signals, and calendar features to forecast demand at the feeder, substation, or zone level. Day-ahead and week-ahead forecasts with confidence intervals feed procurement, dispatch, and switching decisions. Reduces reserve procurement cost by narrowing the uncertainty band around the forecast. For distribution utilities managing renewable intermittency, demand forecasting combined with generation forecasting improves real-time dispatch quality.
Pipeline and grid anomaly detection
Models trained on your SCADA sensor data that learn the normal operating envelope for each pipeline segment or grid zone. Flags pressure, flow, and temperature deviations that fall outside the model's expected range for the current operating conditions. Alerts surface to control room operators with the contributing sensor signals and the time window of the deviation. Detects slow leaks and equipment degradation events days before they become reportable incidents. Calibrated to reduce false positive alert volume for operators.
Field service routing optimisation
AI routing that schedules and sequences field technician work orders by asset failure probability, geographic proximity, crew skill, and vehicle availability. High-risk assets get scheduled first. Travel time and fuel cost are minimised across the day's work list. Re-routes dynamically when emergency callouts arrive. Integrates with your existing work order management system. Reduces technician drive time and increases the number of work orders completed per crew per day without adding headcount.
The optimisation model runs as a mixed-integer programming solver (Google OR-Tools or similar) over the full day's work order set, balancing asset criticality weights against travel distance matrix derived from real road network data. When a priority-1 callout arrives mid-shift, the model re-solves the remaining schedule in under 30 seconds and dispatches re-route instructions via the field mobile app. Crew skill matrices ensure a transformer oil sampling work order is only assigned to a technician certified for high-voltage proximity work. Integration connects to your CMMS (IBM Maximo, SAP PM, or Infor EAM) via REST API to pull open work orders and push completion status without duplicating data entry. Fleet telemetry from vehicle GPS feeds actual position data rather than assumed depot locations for real-time re-routing accuracy.
Renewable energy output forecasting
Short-term and day-ahead output forecasts for solar and wind assets using weather model inputs, historical production data, and equipment performance curves. Forecasts at the asset and portfolio level with uncertainty bounds. Feeds dispatch decisions, grid scheduling submissions, and imbalance cost reduction. For operators managing curtailment decisions, output forecasting at 15-minute resolution improves the accuracy of curtailment scheduling and reduces revenue loss from over-curtailment.
Solar output models combine NWP (Numerical Weather Prediction) irradiance forecasts from ECMWF or NOAA GFS with sky camera imagery (where available) and historical clear-sky index data to estimate plane-of-array irradiance, then apply panel efficiency curves and inverter loss factors to produce AC output forecasts. Wind models use hub-height wind speed and direction forecasts combined with turbine power curves and wake loss corrections for multi-turbine sites. For both asset classes, a gradient boosting model (XGBoost or LightGBM) is trained on 12-24 months of historical production data to correct systematic NWP bias at your specific site location. Forecast horizon: 15-minute resolution out to 6 hours (intra-day), hourly resolution out to 72 hours (day-ahead and two-day-ahead). Forecast uncertainty is quantified as prediction intervals (P10/P50/P90) rather than point estimates, giving your dispatch team the information to make risk-adjusted scheduling decisions.
Customer billing anomaly detection
Models trained on smart meter interval data and customer billing history that detect readings inconsistent with a customer's consumption pattern. Flags potential meter faults, non-technical losses (theft), and billing errors before they accumulate across billing cycles. For large commercial customers, consumption anomalies surface in near real time. For residential customers, anomaly flags trigger a review cycle before the bill is issued. Reduces revenue leakage and reduces the volume of billing dispute calls that reach your contact centre.
The detection model builds a consumption baseline per customer using their 12-24 months of interval meter data, segmented by day type (weekday, weekend, holiday) and season. Isolation Forest is used to identify consumption windows that deviate from the expected distribution for that customer's segment and tariff class, reducing false positives that fire on legitimate consumption spikes (heatwaves, industrial production ramp-ups). Z-score thresholds applied to rolling 7-day and 30-day consumption windows catch slow drift anomalies (meter running fast or slow) that point anomaly detection misses. For smart meters transmitting interval data via AMI (Advanced Metering Infrastructure) over DLMS/COSEM or ANSI C12.22 protocols, near-real-time anomaly detection runs on the data stream within minutes of meter transmission. Anomaly flags are categorized: suspected meter fault (consistently low readings), suspected non-technical loss (reads drop to near-zero on an active commercial account), and billing discrepancy (billed consumption inconsistent with interval data sum). Each category triggers a different downstream action: field inspection, fraud investigation workflow, or billing correction, respectively.
Which asset failure or grid cost problem do you want AI to reduce?
Maintenance surprises, forecast inaccuracy, or pipeline anomalies: tell us the specific operational problem and we will assess which AI system addresses it and what your sensor data supports.
We build AI systems that optimise HVAC and building load using occupancy data, sub-metering, weather inputs, and tariff structure. The model manages pre-conditioning and demand peak shaving to reduce both energy consumption and demand charges. For building operators or energy service companies managing a portfolio of commercial assets, the same model architecture runs across all buildings with asset-specific calibration.
Predictive maintenance models for energy assets use time-series sensor data to detect the early signatures of equipment degradation before failure. For a transformer, the relevant signals include oil temperature, dissolved gas analysis readings, load current, and ambient temperature over time. For a rotating machine such as a turbine or pump, vibration frequency spectra, bearing temperatures, and oil pressure are the primary signals. The model learns the normal operating signature of each asset class and identifies deviations from that baseline that correlate with historical failure events in your maintenance records. Output is a ranked list of assets by current failure probability, the contributing sensor signals, and a recommended inspection or maintenance action. The model operates on a rolling window of sensor data -- typically daily or hourly -- and updates the risk ranking continuously. For assets where failure causes significant outage cost or safety risk, a 2-4 week prediction horizon gives maintenance teams enough time to plan and execute the intervention before failure occurs.
A grid demand forecasting model for distribution-level management needs historical load data at the granularity you want to forecast -- typically hourly or 30-minute interval data at the feeder or substation level, going back 2-3 years. Beyond historical load, the model improves significantly with weather data: temperature is the strongest external driver of electricity demand, but humidity, wind, and solar irradiance are also relevant. Calendar features -- day of week, public holidays, school terms -- capture regular demand patterns that weather alone doesn't explain. For distribution utilities serving industrial customers, industrial production schedules and shift patterns are significant inputs. Output is a day-ahead or week-ahead load forecast by feeder or zone with confidence intervals. The forecasts feed procurement decisions (how much reserve to commit), dispatch scheduling, and network switching decisions. We assess your metering infrastructure and historical load data availability in discovery to determine the achievable forecast granularity and accuracy.
Pipeline anomaly detection uses the continuous sensor readings from your SCADA system -- pressure at multiple points along the pipeline, flow rates, temperature, and valve positions -- to detect deviations from the expected operating envelope. A baseline model learns the normal relationship between these sensor readings under different operating conditions: flow rate, ambient temperature, product type, and pressure profile. When a sensor reading or a combination of readings deviates from the predicted baseline by more than a threshold, an alert is generated. The alert includes the sensor IDs, the magnitude of the deviation, and the time window over which it developed. This is designed to surface two types of events: slow leaks that develop gradually over hours or days (a gradual pressure drop below the model's expected value for the current flow conditions) and rapid events such as a rupture or valve failure. The detection threshold is calibrated to minimise false positives -- reducing alert fatigue for control room operators -- while maintaining sensitivity to genuine anomalies.
Energy consumption optimisation for commercial buildings uses building management system data -- HVAC set points, occupancy sensors, sub-metering data by zone, and external weather -- to reduce energy use while maintaining comfort targets. A model learns the thermal dynamics of the building: how long it takes to cool or heat each zone given the current weather, occupancy, and equipment settings. This allows the model to pre-cool or pre-heat a building during off-peak tariff periods rather than running HVAC at full load during peak tariff hours. For buildings with demand charges (billed on peak 15-minute demand), the model manages load across the building to shave the demand peak. Output is a set of HVAC set point recommendations or, where integration allows, direct set point commands to the BMS. For a commercial building with annual energy costs above USD 200,000, consumption optimisation typically reduces energy cost by 10-20%. We assess your BMS data access and metering infrastructure in discovery.
Work with us
Tell us what you need. We'll tell you what it would take.
We scope AI for Energy and Utilities in 30 minutes. You walk away with a clear cost, timeline, and approach. No commitment required.
Scope and cost agreed before work starts. No surprises. No obligation.
Working prototype within 3 weeks of kickoff.
Pay by milestone. You see progress before each invoice.
60-day post-launch warranty. Bug fixes, UI tweaks, and deployment support. No retainer.
