AI-Powered Virtual Power Plant Trading Platform
Built from the ground up for a leading listed energy storage company, this VPP operations platform combines AI-driven wholesale market forecasting with automated day-ahead bidding — delivering risk-aware bid curve generation and live rolling simulation demonstrated at SNEC 2025 in Shanghai.
Revenue Uplift
~50%
vs. baseline strategy in one-month simulation
Forecasting Modules
5
Price · PV · Load · Weather · RT Market
Debut Exhibition
SNEC 2025
Shanghai · Live rolling simulation
Project Overview
A publicly listed energy storage company sought to launch a Virtual Power Plant (VPP) business line and showcase it at SNEC 2025 — China's premier solar and storage exhibition. We built the entire platform from scratch: a real-time VPP operations dashboard spanning multiple sites, a five-module AI forecasting engine covering day-ahead and real-time wholesale clearing prices, PV generation, grid node loads, and weather, and an automated bidding system using MPC+MILP optimization. For the live exhibition demo, we ingested publicly disclosed Chilean electricity market data and ran a continuous rolling simulation — generating daily bid curve recommendations exactly as the platform would before each real market deadline.
Project Challenges
Wholesale Price Volatility
Electricity clearing prices in day-ahead and real-time markets exhibit strong nonlinearity driven by renewable generation mix, demand swings, and participant strategy — making high-accuracy forecasting a foundational requirement.
Multi-dimensional Forecasting Pipeline
Five interdependent forecast streams — day-ahead prices, real-time clearing prices, PV generation, grid node loads, and weather — must run in concert, each demanding tailored feature engineering and model tuning.
Risk-aware Bidding Under Uncertainty
Converting probabilistic forecasts into concrete market bid curves requires balancing expected revenue against settlement risk, with optimization strategies that adapt to each operator's risk appetite.
Production-grade Live Exhibition
The SNEC 2025 demonstration required the platform to run fully autonomously on a fixed daily schedule — ingesting market data, generating forecasts, and producing bid recommendations in front of live exhibition visitors.
Solution Design
Ensemble AI Forecasting Engine
Stacked ensemble of gradient-boosting models (LightGBM, XGBoost) and deep learning (LSTM) delivers short-horizon wholesale price, PV generation, and load forecasts with quantified uncertainty.
MPC + MILP Bidding Optimizer
Model Predictive Control combined with Mixed-Integer Linear Programming translates daily forecasts and user risk preferences into optimal day-ahead bid curves, submitted automatically before each market deadline.
Multi-site VPP Operations Platform
Unified real-time dashboard for monitoring and dispatching multiple generation and storage sites, with automated bidding workflows, alert management, and full audit trails.
Closed-loop Strategy Validation
Post-hoc analysis compares AI-generated bids, industry-standard baseline strategies, and theoretical oracle outcomes against actual market clearing data — quantifying performance gain and informing ongoing model improvement.
Real-time Rolling Simulation at SNEC 2025
To demonstrate the platform's capabilities live at the exhibition, we ran an end-to-end rolling simulation using publicly disclosed Chilean electricity market data — replaying one month of market days and generating daily bid recommendations on the exact schedule a real operator would follow.
Daily Rolling Forecast & Bid Generation
Each simulated day, the platform ingested market disclosure data up to the day-ahead bidding deadline, ran all five forecasting modules, and automatically generated the recommended bid curve — mirroring real-world VPP operations.
Three-way Strategy Benchmarking
Results were benchmarked against an industry-standard baseline strategy and the theoretical oracle (perfect-foresight) optimal strategy, giving visitors a transparent view of the AI system's relative performance.
Post-hoc Market Clearing Validation
After each simulated day settled, actual market clearing prices were fed back into the system to calculate realized revenue, forecast accuracy metrics, and cumulative performance — validating the platform's real-world readiness.
System Interface
Station Overview
Real-time monitoring of multi-site BESS and PV assets — tracking state of charge, power flows, and grid connection status across all sites on a single dashboard.
Prediction & Forecasting
Ensemble AI forecasts for day-ahead prices, real-time clearing prices, PV generation, and load — with quantified uncertainty bands to support informed bidding decisions.
Revenue Analysis
Track and compare the revenue performance of different VPP bidding strategies (including the benchmark strategy, AI strategy, and theoretical optimal strategy) on a daily basis.
Prediction Accuracy Analysis
Monitor the prediction outcomes and accuracy for various targets on a daily basis to promptly detect significant model drift and execute real-time model calibration or retraining.
Project Results
Have a Similar Project?
Chat with our expert team to get customized solutions and practical recommendations.