Introduction of a thermal peak-shaving strategy for buffer storage utilization in multi-residential buildings and prediction of the dampened power profile by data driven models

Peak power demand poses a major challenge for renewable district heating networks, as it is often met by gas-fired auxiliary boilers. Optimized utilization of decentralized thermal buffer tanks within customer-side hydraulic installations offers a promising approach to contributing in solving this issue. For a multi-unit residential building with parallel space heating and domestic hot water supply and a buffer tank, a peak-shaving control strategy is developed in a simulation tool that leads to 53.3% peak power reduction compared to a typical pump control based on a temperature threshold and 31.9% peak power reduction compared to the direct heat consumption power of a system without a buffer storage. Since fast prediction of expected power and return temperature profiles are important for district heating operators to adapt their operational strategies, linear dynamic and neural network approaches are developed and evaluated. Autoregressive models with exogeneous inputs (ARX) achieved quite good prediction accuracy (5.4% mean percentual error (MAPE) for power and 4.7% for return temperature), while being computationally efficient (11 s on a 32 GB CPU). A Convolutional Neural Network (CNN) – Long Short-Term Memory (LSTM) architecture provides higher accuracy, but is computationally more expensive (40 s on a 32 GB GPU). The best CNN-LSTM model, resulting from a grid search, is evaluated for yearly space heating demands from 74 to 354 MWh and space heating to domestic hot water demand ratios from 0.8 to 8.2. The yearly MAPE for power prediction varies from 2.6 to 3.9% and for return temperature prediction from 0.7 to 1.0%.