| In recent years,the penetration rate of distributed photovoltaic(PV)in distribution networks is increasing.The actual load coupled with the random PV output forms generalized loads with greater uncertainty,which brings severe challenges to the safe and stable operation of distribution networks.Accurate distributed PV output forecasting and load forecasting are the basis to guarantee the safety and economy of distribution network operation.However,at present,most distributed PVs are installed behind the meters and lack proprietary metering devices,which makes PV output invisible to distribution network operators and greatly affects PV output forecasting.In addition,distribution network loads have a bottom-up multi-level structure,but existing load forecasting methods are difficult to meet the aggregation consistency requirement of multi-level loads,which increases the burden of distribution network operation decisionmaking.In this context,this study applies deep learning technologies such as generative adversarial net and long short-term memory neural network to achieve accurate estimation of PV output and multi-level load collaborative forecasting for distribution networks with distributed PV.The specific work is as follows:(1)Aiming at the intermittency and randomness of distributed PV output,a short-term distributed PV output interval forecasting method combining scenario generation and interval forecasting is proposed.Firstly,long and short-term memory neural network is used to make the point forecasting of the day-ahead PV output.On this basis,the conditional generative adversarial net is used to generate PV output day-ahead scenarios.Furthermore,the upper and lower bounds of short-term PV output forecasting intervals under different confidence levels are obtained by the confidence interval expansion method using the scenarios.The calculation results show that compared with the traditional Bootstrap method,the forecasting interval constructed by the proposed method not only has high coverage but also has narrower width,which can provide more reliable boundary parameters for dispatching operation decisions of distribution networks.(2)Aiming at the invisibility of part of distributed PV output,a two-stage identification method of distributed PV output based on missing data reconstruction technology is proposed considering the characteristics of no PV output at night.Considering that the PV output is equal to the actual load minus the net load,firstly,the improved generative adversarial net is used to fill the missing data of the actual load in the daytime,and then the rough identification of PV output is realized.On this basis,the measured output data and installed capacity information of the PV proxy are fully considered to achieve the refined identification of the output and installed capacity of the target PV power station.The calculation results show that the proposed method can accurately identify the invisible distributed PV output by semi-supervised learning,which has great practical application value.(3)Aiming at the aggregation consistency requirement of multi-level load forecasting in distribution networks,a multi-level load collaborative forecasting method based on distributed optimization idea is proposed.Firstly,a multi-level collaborative load forecasting framework adapted to the hierarchical characteristics of distribution networks is constructed using the distributed optimization concept based on the alternating direction method of multipliers.On this basis,a forecasting algorithm combining long short-term memory neural network and federated learning is proposed to realize the integrated accurate forecasting of multi-level load by aggregating load forecasting results from bottom to top.The calculation results show that the proposed method can not only meet the requirement of aggregation consistency of load forecasting results,but also jointly improve the short-term forecasting accuracy of all loads. |
PDF Full Text Request