Load forecasting and optimal configuration of dual-carbon microgrid energy systems under climate change

Achieving carbon peaking and carbon neutrality requires energy systems that can remain economical and low-carbon under long-term climate change. This paper develops a climate-sensitive dual-carbon energy system (DCES) planning framework for an integrated microgrid. A CNN–LSTM–attention forecasting model is first used to generate electricity, cooling, and heating load profiles together with renewable-output information under future climate scenarios. These profiles are then embedded into a microgrid-based integrated energy-system model with bidirectional grid interaction. The planning problem is formulated as a tri-objective optimization task that jointly minimizes annual total cost, life-cycle carbon emissions, and grid-interaction intensity. To solve this constrained and nonlinear problem, a Multi-Objective Auxiliary task-Decomposed Deep Q-Learning algorithm (MOADQL/DP) is proposed, combining decomposition-based multi-objective search, reinforcement-learning-assisted operator selection, and adaptive auxiliary-task feedback. The Shanghai office-building case study shows that the CNN–LSTM–attention model provides more accurate multi-energy load forecasting than the compared baseline models. Incorporating climate-responsive load shifts changes both system configuration and dispatch strategy, reducing total cost and carbon emissions by 6.2% and 4.8%, respectively, compared with the reference case without climate-responsive load modeling. Equipment-efficiency degradation under climate stress increases cost and emissions by 3.1% and 2.4%, while bidirectional grid interaction improves renewable-energy utilization by 5.6%. Carbon-neutrality pathway analysis further indicates that 2050–2060 pathways provide a more balanced trade-off between emission reduction and near-term cost pressure than more aggressive early-neutrality schedules.