Abstract: In response to the serious three-phase imbalance problem caused by the randomness and uncertainty of nonlinear loads in the low-voltage distribution system of the air separation plant, a load balancing optimization method based on an improved RNN model is studied. Using improved particle swarm optimization algorithm to optimize the initial weights of RNN and construct a high-precision short-term nonlinear load forecasting model, providing forward-looking data for subsequent optimization. Combining the three-phase four wire system imbalance compensation method following Steinmetz theory with the load optimization adjustment strategy aimed at reducing imbalance and commutation frequency, a complete "prediction compensation adjustment" closed-loop optimization scheme is formed. Through optimization analysis, it has been proven that this method can significantly improve the three-phase balance of the system, effectively improve the power factor, significantly reduce network losses, and unleash the power supply potential of the system.