Network Security of Distributed Photovoltaic Power Generation Prediction System Based on BiLSTM and AES

Abstract

In the distributed photovoltaic power generation prediction system, the traditional power prediction algorithm has insufficient processing ability for the nonlinear characteristics of distributed photovoltaic data, resulting in low prediction accuracy, and the data transmission process is vulnerable to malicious attacks, with the risk of tampering and leakage, affecting network security. This paper combines AES (Advanced Encryption Standard) technology and BiLSTM (Bidirectional Long Short-Term Memory) machine learning algorithm to design a distributed photovoltaic power prediction system that takes into account both prediction accuracy and network security to improve power prediction accuracy and ensure data security. The study improves data quality through wavelet denoising and normalization processing. The BiLSTM model is used to construct a time series prediction framework, a bidirectional network is used to capture the nonlinearity and complex dependencies in distributed photovoltaic data, and the ReLU (Rectified Linear Unit) activation function is used to optimize the nonlinear feature learning ability. In the data security design, AES-256 encryption technology is used to encrypt data, and GCM (Galois/Counter Mode) is combined to provide efficient encryption and integrity verification. The system's prediction and security module are integrated through a distributed deployment solution to support real-time power prediction and secure data transmission in an edge computing environment. Experimental results show that, compared with ARIMA, LSTM and RNN models, the MSE and MAE of BiLSTM-based prediction results are reduced by 19.4%-35.7% and 17%-31.3% respectively. After combining with AES technology, it is better than GRU+DES, LSTM+RSA and RNN+Blowfish systems in data integrity and encryption efficiency. In actual application test experiments, the system's RMSE (Root Mean Square Error) and MAPE (Mean Absolute Percentage Error) indicators were reduced by 9.4%-17.1% and 11.4%-15.4% compared with traditional models, the encryption throughput was increased by about 15%-23%, and the integrity interception rate reached 100%. The study verified the high predictive ability and security of the system in distributed photovoltaic scenario. It provided reliable technical support for the intelligence and security of photovoltaic power generation systems.