Agentic Workflow for C&I Site Energy Cost Minimization

The single largest line-item saving. By forecasting site load and coordinating battery discharge or load curtailment during predictable 15–30 minute peak windows, the workflow actively flattens the demand profile. This directly reduces the capacity-based portion of the utility bill, which can represent 30–50% of total energy costs for C&I customers.

Automates the capture of intraday price spreads. The system continuously ingests real-time pricing (RTP) or day-ahead forecasts, charging on-site storage during low-price periods and discharging during high-price periods. This turns the battery from a passive backup asset into an active revenue-generating or cost-avoidance tool, improving its payback period.

Maximizes the value of on-site PV generation. Instead of exporting excess solar at low wholesale rates (or facing curtailment), the workflow intelligently shifts flexible loads (e.g., HVAC pre-cooling, EV charging) or charges the battery to consume that energy on-site. This displaces higher-priced grid power and improves the effective $/kWh value of the solar asset.

Eliminates the need for facility managers or energy analysts to manually monitor prices and issue dispatch commands. The autonomous control loop operates 24/7/365, making faster, data-driven decisions than a human team could manage. This reallocates skilled labor from tactical operations to strategic analysis and exception management.

Proactively manages site operations to avoid utility-imposed penalties. By integrating with utility DR/curtailment programs and forecasting grid stress events, the system can automatically shed non-critical load or dispatch storage to fulfill commitments, avoiding costly failure penalties. It also optimizes power factor correction in real-time to avoid poor power factor charges.

Preserves capital investment in batteries. The workflow doesn't just maximize short-term revenue; it uses degradation models to evaluate the long-term cost of each battery cycle. By optimizing for total cost of ownership, it avoids excessive cycling during low-value periods, extending the usable life of the storage asset and protecting its residual value.

This agent ingests real-time and day-ahead pricing feeds from the local utility or ISO (e.g., CAISO, PJM), along with site-specific load forecasts. It models future electricity costs, with a special focus on predicting peak demand windows where demand charges are incurred. The forecast triggers downstream optimization 5-60 minutes ahead of real-time, creating the decision window for storage and load-shaping actions.

A dedicated system agent polls the site's solar PV inverters and Battery Management System (BMS) for real-time generation output, state-of-charge (SoC), and discharge capabilities. It enforces operational guardrails (e.g., minimum SoC for resilience) and provides a live inventory of available, dispatchable energy to the optimization engine. Integration is via Modbus, OPC-UA, or vendor-specific APIs.

This integration agent acts as the secure bridge between the optimization logic and the site's BMS (e.g., Johnson Controls, Siemens, Honeywell). It translates high-level curtailment or setpoint adjustment commands (e.g., slight HVAC temperature float, lighting dimming) into native BMS protocols. It also ingests real-time submeter data to verify load response and manage exception states.

The core decision-making agent. It runs a constrained optimization model every 5-15 minutes, weighing forecasted prices, available storage, flexible load capacity, and degradation costs. Its objective is to minimize the sum of energy charges and demand charges over the billing period. It outputs clear dispatch signals: charge battery, discharge to offset grid pull, or shed X kW of flexible load.

A critical control layer. Before any automated load curtailment action exceeding a pre-defined threshold (e.g., >50 kW) is executed, the workflow pauses and routes a notification for approval to a facility manager via Slack, Teams, or a dedicated dashboard. This ensures operational safety and allows for manual override based on real-time site conditions not captured in the model.

An observability agent continuously tracks key metrics: forecast vs. actual prices, battery cycle efficiency, demand charge avoidance, and command execution success rates. It detects anomalies (e.g., storage not responding) and alerts engineering teams. This dashboard provides the audit trail and ROI reporting essential for scaling the workflow across multiple sites.

The workflow's primary actuator. Agents send real-time dispatch signals (e.g., 'discharge battery at 500kW', 'shed HVAC load') via OPC-UA, BACnet, or Modbus protocols. This integration enables direct control over HVAC setpoints, lighting circuits, and on-site generators, translating price signals into physical load adjustments. A fault-tolerant command layer with heartbeat monitoring is essential to prevent unsafe states.

Agents query real-time State of Charge (SoC), temperature, and power limits, then issue charge/discharge setpoints. Integration must respect manufacturer-specific REST or MQTT APIs (e.g., Tesla Powerhub, Sungrow) and embed degradation-aware cycling logic to protect asset lifespan. This is the core lever for energy arbitrage and demand charge reduction.

Real-time ingestion of 5/15-minute interval meter data (via utility APIs or direct meter reads) provides the baseline load profile. This is combined with day-ahead and real-time Locational Marginal Price (LMP) feeds from ISOs like CAISO or PJM. Agents use this fused data stream to calculate the instantaneous cost of grid power, which triggers storage or curtailment actions.

Agents pull real-time and forecasted solar generation data (via SolarEdge, Enphase, or generic Modbus). This informs the net load calculation and determines when excess solar can charge on-site storage instead of being exported. Integration includes handling cloud-cover-induced ramps, requiring nowcasting inputs for stable control.

The workflow's decision engine (built on frameworks like LangGraph or custom agents) pushes all decisions, sensor readings, and calculated savings to a central historian (e.g., InfluxDB, TimescaleDB) and energy management platform UI. This provides audit trails, performance dashboards, and a manual override interface for facility managers, ensuring human-in-the-loop governance.

For sites enrolled in utility demand response (DR) or ancillary service programs, agents must integrate with program-specific gateways (e.g., OpenADR). This allows the workflow to automatically respond to DR event signals, curtail load, and submit verification data, creating an additional revenue stream while adhering to strict performance penalties.