With the continuous growth of global energy demand, the development of oil and gas resources has gradually shifted to complex oil and gas reservoirs. These reservoirs have the characteristics of complex geological conditions, high development costs and great uncertainty. The traditional production prediction methods are difficult to meet the accuracy requirements, resulting in the increase of development decision-making risk. In order to solve the problems of insufficient yield prediction accuracy and difficult risk quantification of economic evaluation in complex reservoir development, this study proposes a coupling analysis method based on modified Bayesian network (MBN). By introducing the dynamic structure optimization mechanism, a multi-level network model integrating geological parameters, engineering parameters and dynamic data is constructed, and the Markov chain Monte Carlo (MCMC) algorithm is used for parameter learning, which effectively solves the problem of solidification of the traditional Bayesian network structure. Taking an offshore sandstone reservoir as an example, comparative experiments show that the prediction accuracy of MBN model is significantly better than that of traditional Bayesian network, LSTM and ARPS methods, and the confidence interval coverage is increased to 93.2%. In terms of economic evaluation, the economic benefits of different development schemes are quantified by coupling the production prediction results and Monte Carlo economic risk simulation. The well pattern infill+intelligent gas injection scheme increases the net present value (NPV) by 39% and the value at risk(VaR@95%) reduced by 25%, becoming the best development scheme. The research results provide a quantitative analysis tool with both prediction accuracy and economic feasibility for reservoir development decision-making, and have important reference value for the development of complex oil and gas resources.