摘要
短期建筑冷却负荷预测是许多建筑能源管理任务的重要基础。针对传统预测模型将相关性较低的特征作为输入会导致模型降低预测精度,提出基于注意力机制的深度强化学习(深度确定性策略梯度,AM-DDPG)短期建筑冷负荷预测方法。该方法首先将数据进行归一化处理。其次,将预测问题建模为马尔可夫决策过程。其中状态为当前气象数据以及历史冷负荷;动作为下一小时建筑冷负荷预测值;奖赏为下一小时建筑冷负荷真实值与预测值的差。最后通过数据深度强化学习智能体学习并且更新内部网络参数得到预测模型。由于存在部分相关性较低的特征,在深度强化学习智能体的Actor网络加入注意力机制,使模型聚焦于相关性较高的特。实验结果表明,在平均绝对误差评估指标下,AM-DDPG比BP、LSTM、DDPG分别降低了22.12%、10.99%、21.40%;在均方根误差的评估指标下AM-DDPG比BP、LSTM、DDPG分别降低了10.43%、4.58%、7.09%。
Short-term building cooling load forecasting is an important basis for many building energy management tasks.In view of the fact that the traditional prediction model takes low-correlation features as input,which will lead to the reduction of prediction accuracy of the model,an attention mechanism-based deep reinforcement learning(Deep Deterministic Policy Gradient,AM-DDPG)short-term building cooling load prediction method is proposed.The method first normalizes the data.Second,the prediction problem is modeled as a Markov decision process.The state is the current weather data and the historical cooling load.The action is the predicted value of the building cooling load in the next hour.The reward is the difference between the actual value and the predicted value of the building cooling load in the next hour.Again,the attention mechanism is added to the actor network of deep reinforcement learning.Finally,the prediction model is obtained by learning from the data and updating the network parameters.The experimental results show that under the mean absolute error evaluation index,AM-DDPG is 22.12%,10.99%and 21.40%lower than BP,LSTM and DDPG respectively.Under the evaluation index of root mean square error,AM-DDPG is lower than BP,LSTM and DDPG are reduced by 10.43%,4.58%and 7.09%respectively.
作者
陈曦尧
张颖
赵立凡
何坤
陈建平
CHEN Xiyao;ZHANG Ying;ZHAO Lifan;HE Kun;CHEN Jianping(School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009;Jiangsu Provincial Key Laboratory of Intelligent Energy Saving in Buildings,Suzhou University of Science and Technology,Suzhou 215009;School of Architecture and Urban Planning,Suzhou University of Science and Technology,Suzhou 215009)
出处
《计算机与数字工程》
2025年第6期1591-1597,共7页
Computer & Digital Engineering
基金
国家重点研发计划“基于数字孪生健康建筑的智能化家庭主动健康信息采集与应用集成技术平台”(编号:2020YFC2006602)
国家自然科学基金项目“面向大型公共建筑节能的强化学习方法及运行数据框架体系研究”(编号:62072324)
江苏省高校自然科学基金项目“小样本条件下的并行迁移强化学习方法研究”(编号:21KJA520005)资助。
关键词
深度强化学习
注意力机制
冷负荷预测
deep reinforcement learning
attention mechanism
cooling load prediction
