Addressing the challenge of optimizing energy consumption in central air conditioning systems under coupled multi-parameter and dynamic operating conditions, this study proposes an integrated approach combining cooling load forecasting with multi-objective collaborative optimization to enhance system energy efficiency and operational stability. Utilizing sensor data from a factory, a hybrid RF-CNN-LSTM forecasting model is developed. This model leverages random forest (RF) for feature selection, a convolutional neural network (CNN) for extracting local patterns, and a long short-term memory (LSTM) network for capturing temporal dependencies, achieving high-precision prediction of cooling capacity demand. Subsequently, an energy consumption model encompassing chillers, water pumps, and cooling towers is established. This study employs the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to simultaneously optimize three objectives: minimizing the total system energy consumption, maximizing the coefficient of performance (COP) of the chillers, and minimizing the fluctuation of key operating parameters. The experimental results demonstrate that the forecasting model achieves a coefficient of determination (R²) of 0.9625 and a mean absolute error (MAE) of 143.62 kW on the test dataset. After optimization, the total system energy consumption is reduced by an average of 14.53%, reaching 124.58 kW. The chillers contribute 93.94 kW (75.4%) of this energy saving, while the combined energy saving rate for the chilled water pumps and condenser water pumps reaches 37.76%. Despite a 16.40 kW increase in cooling tower power consumption, the overall system achieved significant net energy savings. A reduction of 4.8 °C in the cooling water inlet temperature led to a 13.8% improvement in chiller COP. The multi-objective optimization strategy significantly reduced the temperature fluctuation of chilled water supply and return by 76.48% and 81.79%, respectively, effectively enhancing operational stability. For scenarios involving sudden increases in cooling load, although the fluctuation of the cooling water inlet temperature slightly increases, the response effectively overcomes the hysteresis typically associated with traditional control systems.