Hybrid learning algorithm with pairwise scatter plotting features for utility-scale energy systems

Dispatch strategies of utility-scale photovoltaic (PV) battery systems have mostly relied on model-based approaches with forecasting generations and demands, due to their intermittent nature. When using data-driven method, it is a challenging process to identify the most adequate models with limited computational power. An advanced hybrid dispatch algorithm is proposed, which combines a Neural Network (NN) with optimization techniques. A sequential NN with dense hidden layers and RMSprop optimiser are integrated to fine-tune the NN parameters. Pairwise scatter plotting method is used to intuitively and efficiently identify the underlying connections of inputs. The model effectively balances accuracy and computational efficiency. Sensitivity analysis identified that electricity tariffs are the most influential feature in system operations, while time and date have minimal impacts. The method allowed the model to identify key features effectively, reducing model complexity and computational costs. Case studies found that, the hybrid algorithm outperforms the traditional methods even without PV generation and demand forecasts. The proposed method only requires a fraction of computational power compared to conventional model-based optimisation method, and around 60 % of that using Long Short-Term Memory (LSTM) method. • Hybrid Machine Learning algorithm for utility-scale energy systems. • Dispatch algorithm does not require PV generation and demand forecasts. • Use Pairwise Scatter Plotting to select key features.