A data-driven trading strategy for wind-solar-battery microgrids in the spot electricity market

The integration of renewable energy sources, such aswind and solar power, into microgrids presents challenges owingto their intermittent output and the volatility of spot markets.This paper proposes a multi-energy complementary operationmode for wind-solar-battery microgrids, enabling engagementin day-ahead and ancillary service markets to optimize tradingamid output uncertainties. We develop a machine learning-basedoptimization framework: Gaussian Mixture Models (GMM) withK-means clustering produce representative seasonal scenarios forwind and solar generation to capture stochastic variations; LongShort-Term Memory (LSTM) networks forecast electricity pricesto model temporal dynamics. These inputs inform a MixedInteger Linear Programming (MILP) model that maximizesmicrogrid profitability while ensuring a series of constraintson microgrid operations and electricity market transactions. Anempirical evaluation is conducted using real operational data,meteorological records, and market prices from the Guangdongspot electricity market in China. We employ a comprehensivevalidation framework including out-of-sample evaluation of theLSTM price forecaster, clustering-based scenario generationpreserving real-world renewable variability, and systematic sensitivity analyses across prediction errors among 5% to 10%and imbalance price coefficients among 1.05 to 1.35. The mainfindings reveal: (1) The data-driven operation mode providesfeasible and reliable supply by coordinating energy storage andancillary market procurement, supported by the LSTM modelthat achieves high prediction accuracy with a mean absoluteerror of 0.0093 on a 20% holdout test set; (2) The microgridexhibits pronounced seasonal characteristics, requiring differentiated trading strategies across seasons; (3) The imbalanceprice in the ancillary service market critically shapes microgridprofitability, with profits peaking around an imbalance pricecoefficient of 1.2 and displaying an inverted U-shaped response.