Research On Time Series Data Forecasting For Smart Grid Edge Computing

Smart grid offers considerable flexibility in energy production and distribution.In order to achieve this flexibility,smart grid must predict changes in terms of supply and demand more accurately,provide intelligent control decisions in real time at the aggregation level as well as at the individual component level,and distribute power generate from various energy sources more efficiently.However,with the development of smart grid,the data of smart grid shows exponential growth.It can be predicted that it will be difficult for the future grid cloud to process such large scale data efficiently and in real time,which makes it difficult to predict power data efficiently and fail to provide prediction in real time consequently.In order to solve the above problems,this thesis proposes an edge-intelligenceoriented ultra-short term load forecasting framework.It transfers the ultra-short term load forecasting task in the smart grid from the cloud to the edge smart devices which provide local services,so as to reduce the computing and bandwidth pressure on the cloud.Meanwhile,this thesis extends the research from one-step load forecasting to multi-step load forecasting.The main research works and contributions are described as follows:1.To address the multiplicity and real-time of ultra-short term load forecasting tasks and the computational and bandwidth pressure at the cloud in the future.This thesis proposes an edge intelligence-oriented ultra-short term load forecasting framework based on the edge-cloud collaboration framework.Firstly,it relies on smart devices to collect power data which will be uploaded to the cloud.In such way,it can make full use of computing resources in the cloud,and use Bi-directional Long Short-Term Memory in the cloud for data modeling.Secondly,the trained model in the cloud is distributed to the edge intelligence devices through the network,and the edge intelligence devices use the model and the data collected from the smart meters to perform load prediction.Finally,the prediction results are transmitted to the cloud or to the user.2.With respect to the problem that the encoder of the sequence to sequence model causes excessive information loss when converting long sequences into internal state vectors,which leads to the degradation of model prediction performance,a sequence to sequence multi-step load prediction model based on Bi-directional Long Short-Term Memory is proposed.Firstly,the input sequence is encoded into a lossy internal state vector by the Bi-directional Long Short-Term Memory.Then,the internal state vector is decoded in the decoder of the sequence to sequence model by Long Short-Term Memory to a new sequence,which is the prediction result.This thesis is validated with real load datasets for each of the two experiments.The experimental results demonstrate the feasibility of using Bi-directional Long Short-Term Memory for ultrashort-term load prediction at edge smart devices.It also illustrates the effectiveness of a sequence to sequence multi-step load prediction method based on Bidirectional Long Short-Term Memory is demonstrated.