AI Energy Consumption Forecasting System Development

Development of an AI system for predicting energy consumption

Electricity consumption forecasting is necessary at all levels of the energy system: the facility (industrial plant, building), the distribution network, and the system operator. Forecast accuracy directly impacts the cost of balancing: every 1 percent error costs grid companies millions of rubles per year.

Hierarchy of forecasting problems

Object level:

• Forecasting building/business consumption for peak load management
• BEMS (Building Energy Management System): optimization of HVAC and lighting operation
• Horizon: 15 minutes - 24 hours

Distribution network level:

• Load forecast for transformer substations
• Load balancing between feeders
• Horizon: 1-7 days

System level (SO UES):

• Forecast of consumption by the UES (unified energy systems)
• Planning of dispatch power plants
• Horizon: 1-7 days, quarterly and annual

Key factors

Weather (40-60% of consumption variability):

• Temperature: the main driver. Heating Degree Days (HDD) and Cooling Degree Days (CDD).
• Temperature-Load curve: U-shaped for residential applications (heating + air conditioning)
• Humidity: apparent temperature, affects air conditioning
• Solar radiation: direct influence on building cooling

Time patterns:

• Daily profile: weekdays (office peak), weekends (residential peak)
• Seasonal: summer vs. winter vs. transitional period
• Holidays: industry is at a standstill, the residential sector consumes differently

Structural changes:

• Commissioning of new businesses/shopping centers
• Electrification of transport: EV charging creates new peak profiles
• Heat pumps: consumption increases in winter

Models for different horizons

Very Short-Term (15 min - 4 hours):

• LSTM with a sequence of the last 24-48 hours
• Feature: recent load, weather (actual + forecast)
• Metrics: MAPE < 3%

Short-Term (1-7 days):

import lightgbm as lgb

features = {
    'load_lag_24h': load_yesterday_same_hour,
    'load_lag_168h': load_last_week_same_hour,
    'temp_forecast': temperature_forecast,
    'hdd': max(0, 18 - temp_forecast),  # Heating Degree Days base 18°C
    'cdd': max(0, temp_forecast - 22),  # Cooling Degree Days
    'hour': hour_of_day,
    'dow': day_of_week,
    'is_holiday': holiday_flag,
    'sunrise_hour': astronomical_sunrise,
    'ghi_forecast': global_horizontal_irradiance
}

model = lgb.LGBMRegressor(n_estimators=500)

Medium-Term (month - year):

• Seasonal decomposition + trend model
• Macroeconomic indicators (GDP → industrial production → energy consumption)
• MAPE 3-7%

Consumption Anomaly Detection

Baseline + anomaly:

def detect_consumption_anomaly(actual, predicted, window=168):
    # Normalized residual
    residuals = actual - predicted
    baseline_std = residuals.rolling(window).std()
    z_score = residuals / baseline_std
    return z_score.abs() > 3.0

# High z-score → possible leak, equipment not working properly
# Low z-score → equipment is stopped (holiday, breakdown)

Abnormal consumption → automatic notification to the enterprise energy manager.

Demand Response integration

The forecast allows for automated demand response:

If there is an expected shortage in the network:

1. SO UES announces a price signal in DAM (Day Ahead Market)
2. The object's BEMS receives a signal
3. Automatic: Flexible load shifting (EV charging, heat storage tank heating)
4. Reducing the peak by 10-20%

For industrial consumers (RTE / KOM): DR contract: you agree to reduce the load by X MW when the signal is given → you receive a bonus. The ML system accurately determines the flexible load (can be transferred without damage to production).

Integration with management systems

• SCADA ACS TP: obtaining actual load data in real time
• ASKUE (Automated System of Commercial Electricity Metering): data from metering devices
• BI systems: Power BI / Tableau dashboards for energy managers
• ERP SAP IS-U: integration for energy supply companies

Metrics:

• MAPE of the daily forecast: < 5% for the system operator, < 3% for the object
• Peak Load Accuracy: peak forecast error < 2%
• Cost savings: reduction of off-balance sheet value

Timeframe: Basic short-term forecasting model for a single object: 3-4 weeks. Hierarchical network-level system with anomaly detection and demand response: 3-4 months.