基于改进的CNN-BiGRU供热量预测方法
Heat Supply Prediction Method Based on Improved CNN-BiGRU
张珂 1曹姗姗 1孙春华 1夏国强 1吴向东2
作者信息
- 1. 河北工业大学能源与环境工程学院,天津 300401
- 2. 河北工大科雅能源科技股份有限公司,河北 石家庄 050000
- 折叠
摘要
提出一种融合Dropout机制和贝叶斯(Bayesian)优化的改进的CNN-BiGRU供热量预测模型.利用Dropout机制,降低预测模型的复杂度,减少神经网络陷入局部最优的风险.采用Bayesian优化算法对预测模型超参数进行优化,提高预测模型的预测精度.结合算例,对改进的CNN-BiGRU供热量预测模型的预测效果及训练集样本占比对预测效果的影响进行分析.改进的CNN-LSTM、CNN-GRU、CNN-BiLSTM预测模型的预测值明显偏离实测值,而改进的CNN-BiGRU预测值模型的预测更加贴近实测值.改进的CNN-BiGRU预测模型各项评价指标均优于其他预测模型,训练时间也相对理想.综合考虑,改进的CNN-BiGRU预测模型的预测效果最佳.预测效果随训练集样本占比增大而改善.当训练集样本占比达到70％后,继续增大训练集样本占比,对改善预测效果的帮助不大.为节约时间成本,推荐训练集样本占比为70％.
Abstract
An improved CNN-BiGRU heat sup-ply prediction model that integrates Dropout mechanism and Bayesian optimization is proposed.The Dropout mechanism is used to reduce the complexity of the pre-diction model and reduce the risk of the neural network falling into the local optimum.The Bayesian optimiza-tion algorithm is used to optimize the hyper-parameters of the prediction model and improve the prediction ac-curacy of the prediction model.Combined with an ex-ample,the prediction effect of the improved CNN-BiG-RU heat supply prediction model and the influence of the proportion of training set samples on the prediction effect are analyzed.The predicted values of the im-proved CNN-LSTM,CNN-GRU and CNN-BiLSTM pre-diction models significantly deviate from the measured values,while the predicted value of the improved CNN-BiGRU prediction model is closer to the measured val-ue.The improved CNN-BiGRU prediction model is better than other prediction models in all evaluation in-dexes,and the training time is relatively ideal.All things considered,the improved CNN-BiGRU predic-tion model has the best prediction effect.The predic-tion effect improves with the increase of the proportion of training set samples.When the proportion of training set samples reaches 70％,continuingly increasing the proportion of training set samples does not have much help in improving prediction performance.In order to save time and cost,it is recommended to have a train-ing set sample ratio of 70％.
关键词
供热量预测/双向门控循环神经网络/卷积神经网络/Dropout机制/贝叶斯优化Key words
heat supply prediction/bidirection-al gated recurrent neural network/convolutional neural network/Dropout mechanism/Bayesian optimization引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据