Peak Shaving Algorithm

The primary objective is minimizing the maximum average power drawn over a utility-defined demand interval (typically 15 minutes). The algorithm continuously monitors real-time site load against a configurable demand limit setpoint. When load approaches this threshold, the battery discharges to shave the peak. The logic must account for ratchet clauses, where a single peak can set charges for months, making precision critical.

Algorithms fall into two categories:

• Reactive Control: Triggers discharge based on real-time load exceeding a static threshold. Simple but can miss rapid load spikes.
• Predictive Control: Uses machine learning forecasting of site load and solar generation to pre-emptively charge the battery before predicted peaks. This ensures state of charge (SoC) availability and prevents premature depletion, maximizing demand reduction for complex load profiles.

The algorithm must maintain a delicate SoC reserve specifically for peak events. Discharging too early leaves no energy for the actual peak; holding too much reserve wastes opportunities for energy arbitrage. Advanced algorithms dynamically adjust the SoC target based on time-of-day, day-ahead load forecasts, and the probability of a peak event occurring within the next demand interval.

Dispatch decisions are bounded by physical and contractual limits:

• Battery C-rate: Maximum charge/discharge power relative to capacity.
• Cycle life degradation: Marginal cost of cycling the battery versus the demand charge avoided.
• Export limitations: Utility rules preventing power export to the grid.
• Round-trip efficiency: Energy losses during charge/discharge cycles. The algorithm solves a constrained optimization problem, often using Model Predictive Control (MPC) , to find the economically optimal dispatch schedule.

Behind-the-meter solar photovoltaics introduce variability. A cloud passing over an array can cause a sudden drop in generation, creating an unexpected net load spike. The peak shaving algorithm must fuse real-time solar irradiance data and short-term generation forecasts to anticipate these dips. It pre-emptively reserves battery capacity to compensate for solar intermittency, ensuring the combined site import remains below the demand limit.

The algorithm's logic is fundamentally shaped by the specific utility rate structure. It must model:

• Time-of-Use (TOU) windows: When demand charges apply (e.g., 12 PM–6 PM weekdays).
• Non-coincident peaks: Charges based on the site's maximum demand regardless of system peak.
• Coincident peaks: Charges based on site demand during the grid's top 3-5 system peaks. The algorithm shifts from simple threshold shaving to strategic, calendar-driven dispatch based on tariff rules.

The primary economic driver behind peak shaving. Commercial and industrial utility tariffs include demand charges based on the highest 15-minute average power draw during a billing period. A peak shaving algorithm targets these specific intervals to minimize the ratchet clause that can lock in high charges for months. The algorithm must predict not just total energy consumption but the precise timing and magnitude of non-coincident peaks that may not align with overall grid stress.

The peak shaving algorithm must balance immediate discharge needs against long-term battery health. Key constraints include:

• Depth of Discharge (DoD): Limiting discharge to 80-90% to preserve cycle life
• C-rate limitations: Respecting maximum charge/discharge rates to prevent thermal stress
• SoC reservation: Holding sufficient charge for predicted afternoon peaks while avoiding full saturation
• Calendar fade compensation: Adjusting dispatch aggressiveness as the battery degrades over its 10-15 year lifespan

Predictive accuracy determines peak shaving effectiveness. The algorithm ingests short-term load forecasts generated by machine learning models trained on historical interval meter data, weather inputs, and occupancy schedules. A 15-minute ahead forecast with less than 5% mean absolute percentage error (MAPE) is critical. The algorithm must also account for forecast uncertainty bands, maintaining a reserve buffer to handle unexpected load spikes that exceed predicted values.

While distinct from peak shaving, TOU arbitrage often runs concurrently on the same battery asset. The algorithm must coordinate these competing objectives:

• Peak shaving: Discharge triggered by site load exceeding a kW threshold
• TOU arbitrage: Discharge triggered by high energy prices ($/kWh)
• Conflict resolution: A hierarchical controller typically prioritizes demand charge reduction over energy arbitrage when the battery's power capacity is constrained
• Stacked services: Advanced algorithms co-optimize both revenue streams using multi-objective optimization

The most sophisticated peak shaving algorithms use Model Predictive Control, which solves a constrained optimization problem over a receding time horizon. The MPC framework:

• Maintains a dynamic model of battery efficiency, thermal limits, and degradation
• Solves a mixed-integer linear program (MILP) at each time step to determine optimal charge/discharge power
• Incorporates updated load forecasts and real-time SoC feedback
• Applies only the first control action, then re-optimizes at the next interval This approach handles the non-linear degradation cost and discrete switching constraints that simpler rule-based controllers cannot address.

Peak shaving algorithms must respect interconnection agreements that often prohibit or limit power export to the grid. The controller enforces a zero-export constraint by dynamically limiting battery discharge to never exceed the site's instantaneous load. In jurisdictions allowing export, the algorithm may participate in demand response programs by discharging during utility-called events, but must carefully track export tariffs and net metering rules that affect the economic return of sending power back to the distribution system.