Objective With the increasing penetration of renewable energy, intraday dispatch of provincial power grids faces the dual challenges of growing net load uncertainty and difficulties in real-time decision-making for transmission corridor congestion. New methods are therefore needed for both uncertainty quantification and combinatorial optimization.

Method The applicability boundaries of the same coherent Ising machine (CIM) in two types of power system problems were investigated. On the control side, large-scale topology selection was formulated as a quadratic unconstrained binary optimization (QUBO) problem. The CIM generated low-energy-state samples in parallel through a single annealing run, and candidate pools were selected using the average pairwise Hamming distance (APD) criterion within each pool. On the forecasting side, a conditional restricted Boltzmann machine (CRBM) with a conditional-bias architecture was proposed. Continuous meteorological and historical load features were incorporated as dynamic biases into the energy function to avoid training instability. CIM hardware was also employed as a negative-phase sampler for persistent contrastive divergence (PCD) training. Simulated annealing was used as a baseline for comparison to assess the applicability of CIM-based annealing sampling in the negative phase of probabilistic modeling. In addition, an adaptive triggering strategy driven by three signals—state changes, solution quality degradation, and solution timeliness—was designed to reduce quantum resource consumption.

Result Based on data from the Shandong power grid and the IEEE RTS-96 test system, the CRBM achieved continuous ranked probability scores (CRPSs) ranging from 0.0092 to 0.0225 across three typical days. CIM-based PCD training exhibited weight divergence, indicating that CIM-based annealing sampling was unsuitable for unbiased sampling in probabilistic modeling. On the control side, the candidate pools generated by the CIM exhibited greater diversity than those generated by simulated annealing and greedy construction, while the algorithm achieved substantially lower execution latency. The adaptive triggering strategy reduced the number of quantum invocations by approximately 58.6%, with a difference in decision returns of only 0.04%.

Conclusion The proposed conditional-bias architecture can incorporate high-dimensional continuous conditioning information while retaining a binary visible-layer structure. Comparative experiments using the same CIM hardware across two types of power system problems demonstrated that Boltzmann sampling was well suited to combinatorial optimization tasks targeting sets of low-energy states, whereas its applicability to negative-phase sampling for probabilistic modeling, which relies on unbiased distribution estimation, required careful evaluation. The tolerance of a given task to deviations in the sampling distribution is a key factor in determining the applicability of CIM.