Accurate forecast of energy consumption is essential in building energy management. Owing to the variation of outdoor weather condition among different seasons, year-round historical weather profile is needed to investigate its feature thoroughly. Daily weather profiles in the historical database contain various features, while different architecture of deep neural network (DNN) models may be identified suitable for specific featuring training datasets. In this study, an integrated artificial intelligence-based approach, consisting of feature extraction, evolutionary optimization and adaptive DNN model, is proposed to forecast week-ahead hourly building energy consumption. The DNN is the fundamental forecasting engine of the proposed model. Feature extraction of daily weather profile is accomplished through clustering techniques. Genetic algorithm is adopted to determine the optimal architecture of each DNN sub-model. Namely, each featuring cluster of weather profile, along with corresponding time signature and building energy consumption, is adopted to train one DNN sub-model. Therefore, the structure, activation function and training approach of DNN sub-models are adaptive to diverse featuring datasets in each cluster. To evaluate the effectiveness of the proposed predictive model, it is implemented on a real office building in the United Kingdom. Mean absolute percentage error of the training and testing cases of the proposed predictive model is 2.87% and 6.12%, which has a 24.6% and 11.9% decrease compared to DNN model with a fixed architecture. With the latest weather forecast, the devised adaptive DNN model can provide accurate week-ahead hourly energy consumption prediction for building energy management system.

能源消耗的准确预测对于建筑能源管理至关重要。由于不同季节户外天气条件的变化，需要全年的历史天气资料来全面调查其特征。历史数据库中的每日天气概况包含各种功能，而深层神经网络（DNN）模型的不同体系结构可能会被标识为适合特定功能训练数据集。在这项研究中，提出了一种基于人工智能的集成方法，该方法包括特征提取，进化优化和自适应DNN模型，以预测每周提前的每小时建筑物能耗。DNN是所提出模型的基本预测引擎。每日天气概况的特征提取是通过聚类技术完成的。采用遗传算法确定每个DNN子模型的最佳架构。即，采用每个具有天气概况特征的群集，以及相应的时标和建筑能耗，来训练一个DNN子模型。因此，DNN子模型的结构，激活函数和训练方法适用于每个聚类中不同特征的数据集。为了评估所提出的预测模型的有效性，该模型在英国的真实办公楼中实施。所提出的预测模型的训练和测试案例的平均绝对百分比误差为2.87％和6.12％，与具有固定体系结构的DNN模型相比，减少了24.6％和11.9％。有了最新的天气预报，