As the construction of urban stormwater detention tanks advances, the problem of their energy consumption is becoming more prominent. At present, some of stormwater detention tanks try to add distributed photovoltaic power generation and energy storage systems to realize zero-carbon operation. However, the power generation of distributed photovoltaic is small and random, and the existing prediction methods are difficult to meet the long-term accuracy requirements. Therefore, in response to China’s strategy of Carbon peaking and Carbon neutrality, and with the goal of green and low-carbon operation of stormwater detention tanks, this paper proposes a multi-model collaborative prediction of photovoltaic power generation of stormwater detention tank based on meteorological and temporal characteristics. The algorithm can realize the accurate prediction of photovoltaic power generation of stormwater detention tanks, so as to reasonably and effectively formulate a flexible control strategy for the load of stormwater detention tank, and create conditions for building the infrastructure of “zero-carbon” cities. This method first detects outliers in the original data through DBSCAN. Then, the random forest algorithm is used for reverse feature selection, and the EEMD algorithm is used to decompose the photovoltaic power data into several intrinsic mode functions and residuals. The meteorological characteristics that affect photovoltaic output are effectively combined with their own time series characteristics. A CNN-BiGRU hybrid neural network prediction model is established to predict the decomposed components separately and obtain the prediction of each component. Finally, all the predictions are superimposed to obtain the final prediction of photovoltaic power generation. This paper takes the real power generation data of stormwater detention tank as an example, the experimental results show that the proposed algorithm has higher prediction accuracy than other neural network algorithms.