Time Series Forecasting Based On Feed-forward Neural Networks

The research on time series forecasting is very important,and its application scenarios can be seen everywhere from e-commerce sales forecasting to stock price forecasting.Time series prediction tasks have also been a long-term research focus in the academic community.With the increasing of complexity of the data format,the model has evolved from a statistical methods based on stable series assumptions to machine learning methods that can handle non-stationary data.However,the challenges like feature extraction and the design of loss functions still exist.We aim at resolving feature extraction problem in time series forecasting by combining with the advantages of neural networks.We use temporal convolutional network,attention mechanism and residual structure to design a network structure called Feedforward Sequential Network(FSN)capable of simulating recurrent neural network modeling sequential correlation.This model is designed to combine the recurrent network's ability to model the sequential characteristic,and overcoming the shortcomings of the recurrent neural network such as gradient vanishing,limited training efficiency,and long-term memory decay.We use sequential convolutional network and the sequential attention mechanism to achieve the representation of sequential characteristic,and use residual structure to prevent the model from network degradation.We verify the effectiveness of the model from theories and experiments on standard data sets.The experimental results show that FSN can improve the accuracy of time series forecasting.The sensitivity analysis experiments on input sequences of different lengths prove that FSN can effectively control training efficiency and prevent the problem of reduced training efficiency as the input length increases in the recurrent neural network.In addition,we also optimize the time series prediction model from the perspective of the loss function.The existing commonly used loss functions fail to focus on the morphological learning and delay between the predicted sequence with the real one,which leads to the deviations of the model's prediction results.The fluctuations and its occurrence time are crucial for many practical scenarios.For these two aspects,we combine DTW and TDI metrics to design a new loss function called Multi-Scale DTW with Temporal Distortion Index(MS-DTWI),we compare MS-DTWI with the commonly used time series prediction loss functions,and use experiments to verify the effectiveness of MS-DTWI.At the same time,in order to verify the effects of hyperparameters in the loss function on training,we design sensitivity analysis experiments,which makes the analysis of MS-DTWI more comprehensive.In order to solve the problem of commodity management and sales forecast for large-scale chain enterprises,we design and implement the "Taurus system".The system can not only carry out inventory management,store management and transaction management,but also perform further correlation analysis of commodities.It can excavate more valuable information from product sales data and use historical product sales data and known available time and date features to predict future product sales.The prediction module not only integrates traditional statistical algorithms,but also the above-mentioned time series forecasting algorithm based on the FSN,they are used to enable the system to adapt to the needs of different data volumes.The system's visualization module displays the correlation analysis and sales forecast capabilities,allowing users to have more intuitively experience on system's intelligence.