Photovoltaic Power Forecasting Based On Recurrent Neural Network And Design Of FPGA Accelerator

As the proportion of distribution networks for photovoltaic(PV)power generation rises,the large-scale integration of PV systems into power grids brings great challenges to production and operation.Research on predictive technologies for PV power generation is urgently required.Accurate forecast results in PV output would suggest electric power institutions responding timely to optimize power generation plans and improve the capacity of the peak flow reduction for gird,as well as arrange maintenance properly to reduce unit loss.The technology of PV forecasting has high economic value for grid operation.This paper introduced the forecasting methods of PV power generation and summarized the currenting technologies.Artificial neural network can solve complex nonlinear problems,so it was used to establish forecasting models.The correlation analysis of the meteorological factors affecting photovoltaic power generation was performed to determine the input data of the model.Complicated weather conditions lead to the intermittent,random,and volatile of PV system,which makes forecasting difficult.Recurrent neural network(RNN)is considered as an effective tool for time-series forecasting.However,when the weather changes intensely,the long-term sequence of multivariate may cause gradient vanishing(exploding)during the training of RNN,leading the results to local optimum.To avoid the above problems and optimize the performance,the RNN with Long Short-term Memory(LSTM)unit was used to establish the model.Furthermore,general-purpose processors like CPU and GPU cannot implement LSTM efficiently due to the overwhelming structure of LSTM unit.In this research,an FPGA-based accelerator was designd for LSTM that optimizes both computation performance and power consumption.The main research contents are as follows:(1)Correlation coefficient was calculated between meteorological factors and solar irradiance for determine the input data.The short-term solar irradiance forecasting models based on BPNN,RBFNN,and RNN was established and evaluated under complicated weather condition.(2)An LSTM model based on the deep structure of RNN was established to compare the forecasting performance of deep learning with other methods.Cross-regional research was designed to prove the scalability of the method.(3)FPGA was used as the implementation of NN acceleration.According to the LSTM forward operation logic,the corresponding hardware modules were designed.The pipeline method of FPGA was used to accelerate the operation in parallel.The FPGA-based LSTM accelerator was compared with the CPU's to test the computing throughput and power consumption.