贝叶斯声誉与扩散策略势博弈驱动的跨园区算力-电力双边匹配OA
Cross-park computing-electricity bilateral matching driven by Bayesian reputation and diffusion-policy potential game
"东数西算"格局下,跨区算力-电力交易在多个自治园区之间的组织面临多重相互交织的障碍,需构建一种可同时内生信任演化、保证匹配稳定并高效求解网络化博弈的跨园区交易组织机制.建立跨园区算力-电力双边匹配市场模型,将卖方履约可靠度刻画为不可观测的隐变量并与其绿色禀赋耦合,构造计及价格、时延、碳强度与声誉四维的成交综合效用.设计贝叶斯声誉演化机制:以Beta后验递推更新卖方声誉、以记忆保持因子实现历史证据的指数衰减,证明声誉得分几乎必然收敛至真实履约能力且后验方差以有效样本量的倒数速率收缩,并据此构造声誉加权的Gale-Shapley延迟接受双边匹配拍卖,证明其匹配结果不存在阻塞对.将卖方在互联网络上的报价激进度博弈建模为加权势博弈,以选择强度自适应的Fermi动力学求解,证明势函数沿演化轨迹单调上升并收敛至网络纳什均衡.提出图神经网络势博弈求解器与扩散策略生成式决策方法.结果显示,基于8买方-6卖方、24时段的跨区算例,声誉机制的平均稳态估计偏差为0.102,其中成交充分园区的偏差介于0.071~0.081,可靠还原卖方真实可信度.相较仅价格匹配基线,社会福利由0.410升至0.756(提升84.4%),平均碳强度由0.303降至0.272 t/(PFlops·h)(下降10.5%),而出清均价仅由517 元/(PFlops·h)变为522 元/(PFlops·h),表明增益源于对优质低碳卖方的精准甄别而非抬价;进一步计入网络化势博弈的完整机制将社会福利推升至0.880,其在社会福利、减排总量、抗操纵性、匹配公平与时延满足5个归一化维度上的性能包络完全包裹仅价格匹配、计及时延与仅计声誉3种配置,相对仅价格匹配基线实现帕累托改进.势函数沿Fermi动力学自341.9单调升至421.0;图神经网络求解器约10步即达95%收敛,较逐点Fermi迭代(约20步)实现约50%的收敛加速;园区数由6扩展至96时,其求解耗时的双对数斜率约1.15(近线性),而Fermi迭代约1.9(近平方),二者耗时差距扩大逾一个数量级.扩散策略的决策分布标准差达0.317,远高于确定性策略的0.020,呈显著双峰结构.鲁棒性方面:恶意虚报卖方占比由0升至50%时,上述机制的社会福利由0.822降至0.741,衰减仅9.8%,而朴素均值声誉由0.791降至0.358,衰减达54.8%;卖方新能源渗透率由30%升至90%时,机制带来的单位算力碳强度降幅由9.7%单调增至17.3%.故将信任的贝叶斯演化内生于匹配环节、将卖方策略耦合纳入势博弈框架、并以图学习与生成式决策加速均衡求解,可在不抬高出清价格的前提下同步改善跨园区算力-电力交易的社会福利、低碳性与抗操纵性;且该机制的减碳价值随电源侧清洁化程度的提高而放大、其求解优势随市场规模的扩大而放大,与"双碳"目标下的电源结构演进形成同向强化.所提机制与面向单园区的3层Stackelberg-演化-交替方向乘子法协同机制在研究对象、机制设计与求解方法上互为补充,共同构成算力-电力协同的完整方法体系.
Cross-region computing-electricity trading among autonomous parks under China's"Eastern Data,Western Computing"initiative is obstructed by many coupled difficulties.A cross-park trading organization mechanism is demanded that endogenizes trust evolution,ensures matching stability,and efficiently solves the induced networked game.A cross-park computing-electricity bilateral matching market was established in which the sellers'fulfillment reliability was characterized as an unobservable latent variable coupled with their green endowment,and a composite transaction utility spanning price,latency,carbon intensity,and reputation was defined.A Bayesian reputation evolution mechanism was designed,where sellers'reputations were updated by Beta-posterior recursion,and a memory retention factor was introduced to achieve the exponential decay of historical evidence.It was proved that the reputation score converged almost surely to the actual fulfillment capability,and the posterior variance contracted at a rate inversely proportional to the effective sample size.Based on this,a reputation-weighted Gale-Shapley deferred-acceptance bilateral matching auction was constructed,proving that the matching outcome had no blocking pairs.The sellers'pricing-aggressiveness game over the interconnection network was modeled as a weighted potential game,which was solved by Fermi dynamics with adaptive selection intensity.The potential function was proved to increase monotonically along the evolutionary trajectory and converge to a network Nash equilibrium.A graph neural network(GNN)-based potential game solver and a diffusion-policy generative decision-making method were proposed.The results showed that,in an 8-buyer/6-seller,24-period cross-region case,the mean steady-state estimation error of the reputation mechanism was 0.102,and the error ranged from 0.071 to 0.081 for the parks with sufficient trading frequency.This indicated that the mechanism reliably recovered sellers'true trustworthiness.Compared with the price-only matching baseline,social welfare rose from 0.410 to 0.756(+84.4%)and the average carbon intensity fell from 0.303 to 0.272 t/(PFlops·h)(-10.5%),whereas the clearing price shifted only from 517 to 522 yuan/(PFlops·h).It indicated that the gain was mainly attributed to accurate identification of high-quality,low-carbon sellers rather than from price increases.Adding the networked potential game lifted welfare further to 0.880,with its performance envelope over the five normalized dimensions—social welfare,total emission reduction,manipulation resistance,matching fairness,and latency satisfaction—fully enclosing those of the price-only matching,latency-aware,and reputation-only configurations,thereby achieving a Pareto improvement over the price-only matching baseline.The potential function rose monotonically from 341.9 to 421.0 under Fermi dynamics.The GNN solver reached 95%convergence within about 10 iterations,realizing about 50%convergence acceleration compared with the pointwise Fermi iteration requiring approximately 20 iterations.As the number of parks increased from 6 to 96,its solution time scaled with a log-log slope of about 1.15(nearly linear)versus about 1.9(nearly quadratic)for Fermi,and the difference in solution time widened to more than one order of magnitude.The diffusion policy produced a decision distribution standard deviation of 0.317,substantially higher than the 0.020 obtained by the deterministic policy,exhibiting a pronounced bimodal structure.In terms of robustness,as the share of maliciously misreporting sellers increased from 0 to 50%,the social welfare achieved by the proposed mechanism decreased from 0.822 to 0.741,representing only a 9.8%reduction.In contrast,the naive mean reputation method experienced a decrease from 0.791 to 0.358,corresponding to a 54.8%reduction.As the renewable energy penetration rate of sellers increased from 30%to 90%,the reduction in carbon intensity per unit of computing power achieved by the proposed mechanism increased monotonically from 9.7%to 17.3%.Endogenizing Bayesian trust within matching,incorporating strategic coupling into a potential game framework,and accelerating equilibrium solving through graph learning and generative decision-making jointly improve the social welfare,low-carbon performance,and manipulation resistance of cross-park computing-electricity trading without raising clearing prices.Moreover,the carbon reduction value of the mechanism is amplified as the power generation mixes become cleaner,and its solution advantages become more prominent as markets grow larger,forming a synergistic reinforcement with the evolution of the power generation mixes under the"dual carbon"goals.The proposed mechanism complements the three-level Stackelberg-evolution-alternating direction method of multipliers coordination mechanism for a single park in terms of research objects,mechanism design,and solution methods,together forming a complete methodological framework for computing-electricity coordination.
程乐峰;余风;孙润宝;邹涛
广州大学 机械与电气工程学院,广州 510006广州大学 机械与电气工程学院,广州 510006广州大学 机械与电气工程学院,广州 510006广州大学 机械与电气工程学院,广州 510006
能源科技
跨园区算力-电力交易双边匹配贝叶斯声誉势博弈扩散策略图神经网络算电协同
cross-park computing-electricity tradingbilateral matchingBayesian reputationpotential gamediffusion policygraph neural networkcomputing-electricity coordination
《综合智慧能源》 2026 (9)
68-83,16
国家自然科学基金项目(52171331)广东省普通高校创新团队项目(自然科学)(2024KCXTD031)广州市教育局高校科研项目(2024312278) National Natural Science Foundation of China(52171331)Innovation Team Project for General Universities in Guangdong Province(2024KCXTD031)University Research Project of Guangzhou Education Bureau(2024312278)
评论