| The integrated community energy system is an important carrier for future energy development.It can multi-directionally couple the supply of energy with different attributes,dispatch customers’energy needs in a three-dimensional manner,absorb renewable energy nearby,participates in the grid peak reduction and valley filling,and has the strong self-regulation ability.The core of the integrated community energy system is optimized dispatch,and short-term forecasts on the supply and demand sides are the basis and guarantee for the optimization of dispatch.Precise energy supply and demand forecasting are a necessary prerequisite for the planning,design,management and scheduling of the integrated community energy system,and is the key to the stable and efficient operation of the system,and has great research significance.This paper takes the energy system of the National Speed Skating Stadium as the background to study the integrated community energy system,with the goal of improving the accuracy of forecasting,to study the short-term forecasts on the supply side and the demand side respectively.The characteristics of energy data are analyzed in detail,and the preprocessing methods of energy data are introduced.The power generation of the distributed photovoltaic system on the energy supply side is very sensitive to environmental changes and has the characteristics of randomness and intermittent.This paper proposes a short-term photovoltaic power prediction model based on Stacking ensemble learning method.Firstly,consider the differences in the training principles of different algorithms,perform feature contribution analysis,and build a Stacking integrated learning prediction model.Secondly,use the data provided by the DCS system of photovoltaic power plants to perform iterative training on the single model and the Stacking ensemble model to predict the short-term future photovoltaic power.In two typical day conditions,sunny and rainy,the Cat Boost with the most superior performance in the single model is selected to compare with Stacking.The prediction results show that the RMSE,MAE and R~2evaluation indicators of the Stacking ensemble model are better than the Cat Boost algorithm.Aiming at the problem of strong randomness of demand side power load data and low latitude,this paper proposes a short-term power load multi-step forecasting method based on long-and short-term time-series network(LSTNet).Firstly,the sliding window method is used to construct historical load data into continuous feature maps as input.Secondly,CNN and LSTM are used to capture temporal short-term local information and long-term related information,respectively,and autoregressive models are used as linear components.Then,use the walk-forward verification to guide parameter training.Finally,according to the proposed algorithm model,experiments are carried out on the open data set of UCI Machine Learning Repository,and the power load forecasting of the next week is selected for comprehensive comparison and analysis with the three most popular power load forecasting methods.The experimental results show that the MAPE and RMSE of the prediction model proposed in the article are smaller than other algorithms,can handle the time series relationship between power load data well,and have high prediction accuracy.The two short-term forecasting algorithm models proposed in this paper have high forecasting accuracy,and can provide reference and data support for the optimal dispatch of the integrated community energy system. |
