A Study on the Optimization of Real Time Urban Pumping Station Operation Based on Prediction

As the frequency and intensity of short-duration, intense rainfall increase due to climate change, the risk of urban flooding in areas adjacent to rivers is growing. In particular, urban underground spaces (such as underpasses, underground parking garages, and semi-basement residential units) are structurally vulnerable to water accumulation, and sudden flooding in these areas can lead directly to loss of life. Recent recurring incidents of flooding in underground spaces further highlight the need for proactive drainage management. Drainage pump stations are critical flood prevention infrastructure in low-lying urban watersheds; however, existing rule-based operation methods have limitations in that they respond reactively only after water levels reach a threshold, making it difficult to effectively cope with rapidly changing rainfall patterns. This study proposes an integrated prediction-operation framework that enables real-time, proactive pump operation. Using the Huff time distribution method, a total of 900 rainfall scenarios—each with a total rainfall of 250 mm and a duration of 6 hours—were generated. These were then used as input data in the EPA-SWMM model to simulate runoff. We trained a Bi-LSTM model using the results of the 900 rainfall and runoff simulations to develop an inflow prediction model, and then developed an AI model that operates pumps based on the predicted inflow for the next 8 hours. To analyze performance under extreme conditions, scenarios were simulated where inflow reached 70% and 100% of the total drainage capacity (2,232 m³/s). The analysis compared the current rule-based pump station operation with the AI-based pump station operation developed in this study. The results showed that, unlike the rule-based method which responds reactively after the water level threshold is reached, the AI-based method effectively suppressed peak water levels by proactively operating pumps before the water level rose, utilizing the 8-hour inflow forecast. Furthermore, the maximum water level was reduced by an average of 0.61 m (16.8%) compared to rule-based operation. Notably, under extreme conditions with 100% discharge capacity, a maximum reduction rate of 25.6% was achieved, confirming that the effectiveness of the AI system is further enhanced during extreme rainfall events.