In the logistics management activities of the enterprise, inventory management optimization occupies an extremely important position. Accurately predicting inventory demand has become the foundation and prerequisite of enterprise inventory management, directly impacting profitability and customer service levels. Against the backdrop of explosive growth in global big data, enterprises have access to an increasing amount of information. This calls for the development of more scientific and comprehensive forecasting models that can utilize various complex data, including historical sales data, to achieve more efficient and accurate forecasting objectives. Machine learning models have superior capabilities in handling complex and massive data compared to traditional forecasting models. Therefore, incorporating advanced machine learning theories into inventory demand forecasting can lead to superior forecasting results. This paper is based on the LSTM model for inventory demand prediction. The article suggests a better LSTM forecasting model to fix problems including overfitting, disappearing gradients, and model collapse that were found in earlier experimental investigations. The article utilizes the Dropout mechanism and adds L2 regularization for model structure selection. In the selection of activation functions, the paper proposes the MReLU function, which can enhance the model’s predictive performance and its applicability.