Recent findings reveal that a simple one-layer linear model can significantly outperform transformer-based models. To further enhance time series forecasting performance, we introduce a feedforward neural network (FNN) -based deep learning model, which surpassed most existing models. The novel FNN-based model presents three innovations: input and output structure, FNN division and combination, and an FNN sharing mechanism. The model utilizes three inputs: historical values, historical timestamps, and future timestamps, in order to predict the future demand as its output. FNN division means different inputs corresponding to different FNNs, and the outputs of these FNNs are combined to form the final prediction. The FNN sharing mechanism involves mapping each column (or row) of the matrix inputs to the column (or row) using a common FNN, termed FNN kernel. This FNN neural network framework reduces computational complexity and the number of weights significantly while still providing interpretability and strong performance. Our comparative experiments demonstrate that all three models outperform existing models on two real-world datasets.