Multi-Agent Workflow for EV Fleet Charging Coordinated with Solar Generation

The workflow's primary agent continuously forecasts solar generation and fleet departure schedules. It orchestrates charging sessions to flatten the site's grid power draw, specifically avoiding spikes during peak tariff windows. By shifting load to coincide with solar surplus, it directly reduces monthly demand charges—often 30-50% of a commercial electricity bill—delivering a clear, recurring ROI.

Instead of exporting excess solar energy at low wholesale rates (or facing curtailment), the workflow's dispatch agent uses the EV fleet as a flexible, on-site battery. It dynamically routes surplus generation to charging vehicles, increasing the effective value of each kilowatt-hour produced. This directly improves the economics of the solar asset and enhances sustainability reporting by ensuring clean energy is consumed locally.

Eliminates the manual, daily task of reviewing solar forecasts and manually scheduling chargers or moving vehicles. Fleet managers and facility operators are freed from this reactive coordination work. The workflow handles exceptions (e.g., an unscheduled vehicle needing a rapid charge) through predefined rules and human-in-the-loop escalation, turning a constant operational headache into a managed, automated process.

The scheduling agent treats fleet departure times and minimum state-of-charge requirements as hard constraints. It back-schedules necessary charging from solar and off-peak grid power to guarantee vehicles are mission-ready, eliminating range anxiety and last-minute scrambling. This transforms energy cost optimization from a risky manual guess into a reliable, automated assurance layer for core operations.

Built on orchestration frameworks like LangGraph, the workflow is designed as a modular multi-agent system. This allows for seamless integration of new data sources (e.g., real-time grid carbon intensity), additional asset types (e.g., stationary batteries), or participation in demand response programs. The initial build for cost avoidance creates the architectural foundation for future revenue generation through grid services.

Every decision—kWh sourced from solar vs. grid, carbon avoided, and demand charge savings—is logged with a full chain of reasoning. This creates an immutable, granular record for ESG reporting, regulatory compliance, and stakeholder communication. The workflow turns operational data into defensible sustainability metrics, providing tangible proof of decarbonization progress.

This foundational agent ingests hyperlocal weather forecasts, historical PV performance, and real-time inverter telemetry to predict solar generation for the next 24-48 hours. Simultaneously, it analyzes fleet telematics and scheduled routes to model the fleet's energy demand profile. The accuracy of this dual forecasting directly determines the potential for demand charge avoidance and solar self-consumption optimization.

The core decision-making agent. It runs a constrained optimization model that considers predicted solar output, time-of-use electricity rates, fleet departure times, and each vehicle's state-of-charge (SoC). It outputs a minute-by-minute charging schedule for each charger, dynamically allocating available solar surplus and avoiding simultaneous high-power draws that trigger peak demand charges. Integrates via OCPP 2.0 or manufacturer APIs (ChargePoint, ABB, Siemens) to issue start/stop commands.

A dedicated integration agent that maintains a real-time sync between the charging orchestration layer and the enterprise Fleet Management System (FMS) like Geotab, Samsara, or Verizon Connect. It polls for trip schedules, vehicle location, and current SoC, and pushes back planned charging completion times. This ensures charging schedules never conflict with operational readiness, preserving the fleet's primary utility.

Monitors for deviations: sudden cloud cover, charger faults, or unscheduled vehicle use. When real-time data drifts beyond forecast thresholds, this agent recalculates the schedule or, for major disruptions, routes an alert with recommended actions to a fleet manager via Slack, Teams, or a custom dashboard. This control layer is critical for operational trust and handling edge cases without system failure.

Continuously calculates and logs key performance indicators (KPIs): solar self-consumption percentage, demand charge savings vs. baseline, total grid energy purchased, and carbon offset. It generates daily and weekly reports, attributing savings to specific decisions. This agent provides the ROI justification and operational transparency required for scaling the system across multiple depots.

Implementation typically uses a central orchestrator (e.g., a LangGraph or Temporal workflow) hosted on cloud or on-premise, with lightweight agents interfacing with external APIs. Charger commands are sent via a secure, low-latency message queue (e.g., RabbitMQ, AWS IoT Core). The entire system requires robust API key management, idempotent command handling, and a digital twin simulation environment for testing schedule logic against historical data before live deployment.

The workflow's primary input is a high-confidence, intraday solar generation forecast. This requires integrating with numerical weather prediction (NWP) services (e.g., Solcast, OpenWeatherMap) and potentially on-site pyranometer telemetry via Modbus or OPC-UA. Agents must poll these sources, apply site-specific degradation factors, and output a probabilistic generation curve (e.g., P50, P90) for the scheduling horizon. Data quality checks and fallback to historical baselines are essential for resilience.

Direct control of charging sessions is executed through the Electric Vehicle Supply Equipment (EVSE) management platform's API (e.g., ChargePoint, Enel X, custom OCPP 2.0.1 server). The orchestration layer must authenticate, send remote start/stop commands, and poll for real-time session data (power draw, state of charge). Implementation must handle network latency, command acknowledgment timeouts, and reconcile scheduled vs. actual energy consumption to maintain accurate solar self-consumption calculations.

To schedule effectively, the system needs vehicle availability and battery state. This involves integrating with Fleet Management System (FMS) APIs (e.g., Samsara, Geotab) or direct vehicle telematics to ingest: planned departure times, current location, and initial State of Charge (SoC). This data allows the scheduling agent to prioritize vehicles with imminent departures and calculate required energy, forming the constraints for the solar-optimization problem.

Real-time measurement of site import/export power is non-negotiable for closed-loop control. Integration typically involves Modbus TCP, DNP3, or MQTT connections to a main utility meter or site SCADA system. This data feed provides the ground truth for grid consumption. The control agent uses it as a feedback signal to dynamically adjust charging power, ensuring the site load never exceeds a set threshold (e.g., to avoid demand charges) and maximizes solar self-consumption.

The core logic is implemented in an orchestration framework like LangGraph or Temporal, which manages the stateful workflow between forecasting, scheduling, and control agents. This engine must integrate with observability platforms (e.g., Datadog, Grafana) via APIs to emit custom metrics (solar utilization rate, demand charge savings) and structured logs. All agent decisions, API calls, and exceptions must be logged to an immutable audit trail for performance analysis and regulatory compliance.

A critical integration is with a human-in-the-loop console, often a custom web dashboard or integration with Slack/Microsoft Teams via webhook. This system must push alerts for critical failures (API downtime, charger fault) and present override controls for fleet managers. Scheduled plans and real-time energy flows should be visible, with the ability to manually prioritize a vehicle or pause the optimization, ensuring operational control is never fully ceded.