基于粒子群改进BP神经网络的组合预测模型及其应用

Combined forecasting model based on BP improved by PSO and its application

针对应用广泛的传统人工智能预测BP(Back propagation)神经网络自身局限以及其在处理中长期复杂预测问题中需要样本数量大、泛化能力弱等不足,提出利用粒子群算法优化BP神经网络的学习算法,在此基础上,利用灰色预测方法和自回归移动平均模型(ARIMA)时序预测对历史数据进行初步预测,对中长期预测中数据趋势项和随机项进行模拟;将初步预测的结果作为改进BP神经网络的输入,在此基础上进行训练和预测,构建基于改进BP网络的组合预测模型。以我国1978-2007年能源需求数据为样本,进行实例分析。结果表明:组合预测模型预测精度较BP神经网络、灰色预测方法和ARIMA预测方法分别提高4.8%,6.1%和5.3%,验证了组合预测方法在中长期预测问题处理中的有效性。

To solve the limitations of BP neural network in dealing with complex and long-term prediction, which needs many samples to calculate and its generalization ability is poor, the learning algorithm of BP neural network optimized by PSO （Particle swarm algorithm） was proposed. On the basis of this algorithm, the grey prediction and the ARIMA （Autoregressive integrated moving average model） time series forecasting methods were used to make a preliminary forecast for historical data. The results of initial forecasts were put as the input of improved BP neural network to be forecasted and trained. An improved BP network-based combining forecasting model was built. China＇s energy demands data in 1978-2007 were taken as the sample to be analyzed. The results show that the learning algorithm of BP neural network optimized by PSO has better effects in simulation of the trend data and the random data in medium- and longterm forecasting. The prediction accuracy of combining forecasting model increases by 4.8%, 6.1% and 5.3% respectively compared with those of BP neural network, gray forecasting methods and ARIMA forecasting, which proves the effectiveness of combining forecasting method in the medium-term and long-term forecasting.