Deep learning prediction method based on combined framework of plurality of convolution neural networks

Abstract

The invention belongs to the technical field of data identification and data representation and discloses a deep learning prediction method based on the combined framework of a plurality of convolution neural networks. According to the invention, the method comprises the steps of carrying out the dimension transformation of data; preparing a training data set, a verification data set and a test data set; inputting the data subjected to dimension transformation correspondingly into different convolution neural networks according to periodic state numbers, rearranging the data processed by the convolution neural networks according to the time sequence and then inputting the data into a deep full-connection neural network to obtain a final result; performing early termination training by using the verification data set to obtain a model; predicting the test set to obtain a prediction result. According to the invention, the data of different periodic states are respectively processed through the plurality of convolution neural networks, so that the data rule information is mined in a targeted mode. The number of network layers is reduced. The rearrangement sequence of data after the treatment of convolution neural networks is carried out and then the data are input into the deep full-connection neural network. The data dimension is greatly reduced. The over-fitting is effectively relieved, and the prediction accuracy is improved.

Description

technical field [0001] The invention belongs to the technical field of data identification and data representation, and in particular relates to a deep learning prediction method based on a combination of multiple convolutional neural networks and architectures. Background technique [0002] At present, the existing technologies commonly used in the industry are as follows: The prediction of cell traffic is an important step to alleviate the pressure on the cell. Predicting the change of traffic in a cell in advance is of great significance for resource pre-allocation and load balancing. Through resource planning in advance, not only the pressure on the base station side can be reduced, but the user The user experience will also be greatly improved. Cell traffic forecasting is a time series forecasting problem. Time series is a set of numerical sequences in chronological order, which is a statistical indicator reflecting a certain phenomenon. Time series forecasting problem...

AI Technical Summary

Problems solved by technology

In the case of a large amount of available information, multiple synchronous time series can be obtained, that is, multi-dimensional time series. For example, not only the traffic data of the cell can be used in the problem of cell traffic forecasting, but also the data of the number of users in the cell can also be used for modeling. Key information, for this kind of multi-dimensional time series data, the traditional time series analysis method cannot effectively use multi-dimensional information, resulting in the model not being able to reflect the law of the data well. Moreover, due to some hardware limitations at the base station, the data storage scale is limited, and the data that can be used for training is not long enough, so the time series analysis method is difficult to obtain satisfactory results

Traditional statistical learning methods require manual construction of features. This process requires a lot of business-related knowledge and experience in feature engineering. The process of manually extracting features will inevitably lead to omissions. Inexperienced engineers are likely to spend a lot of time and fail. get better results

Similarly, in the case of less training data and more noise, traditional deep learning methods, such as LSTM, deep fully connected neural network, etc., use the same network to fit data at any time when predicting time series problems. There are different rules in the data, shallow neural network expressiveness is poor, and it is impossible to accurately describe the change of data. Although increasing the number of network layers can present stronger expressiveness for complex data, the parameters in the network also change with the layers. Too many parameters means a large demand for data volume, which may lead to overfitting when the data volume is small, especially when the data volume is small and the noise is large, the convergence of the network will be extremely unstable , making it difficult to obtain stable and excellent prediction results

[0003] In summary, the problems in the prior art are: Traditional deep learning methods have unstable results when the amount of training data is insufficient, resulting in inaccurate prediction results

[0004] The difficulty and significance of solving the above technical problems: It is difficult to express the distribution of data using a simple network structure. Using a complex network structure will lead to an increase in the number of parameters. It is difficult to obtain stable and excellent results when the amount of data is small, and it is difficult to obtain excellent and stable results when the amount of data is insufficient.

Images