To enhance the accuracy of photovoltaic power level classification to meet industry requirements， a classification model based on a Res-MobileCom parallel network was proposed. In data preprocessing， denormalized bilinear interpolation was used to maximize the preservation of data features. Subsequently， through parallel training of simplified residual network（ResNet） and efficient convolutional neural networks for mobile vision（MobileNet）， their outputs were jointly fed into the channel estimation-signal detection network（ComNet） for further data feature extraction， ultimately obtaining the classification results. The experimental results demonstrated that compared to common deep learning models， the Res-MobileCom model retained the feature extraction capability and lightweight nature of ResNet and MobileNet， exhibiting good balance and generalization ability. By using the denormalized bilinear interpolation method and the ComNet for further data feature extraction， the model accuracy improved by more than 10 percentage points， providing a novel approach and idea for improving the accuracy of photovoltaic power level classification models. Future work will focus on stability optimization， cross-task validation， and engineering deployment.

Improving the accuracy of wind power prediction is crucial for ensuring safe and stable operation of the power grid. However， wind power exhibits high randomness and volatility， and traditional prediction methods have limitations in feature extraction and modeling capabilities. Therefore，CEEMDAN-DBO-VMD-TCN-BiGRU， a short-term wind power prediction model integrating complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN）， dung beetle optimizer（DBO）algorithm， variational mode decomposition（VMD）， temporal convolutional network（TCN）， and bidirectional gated recurrent unit（BiGRU） was proposed. CEEMDAN was used to decompose the original wind power data， extracting intrinsic mode functions（IMFs） to capture key features of the time series. The IMFs were divided into high-frequency， medium-frequency， and low-frequency components using sample entropy and K-means clustering. The high-frequency components were selected for secondary decomposition using DBO-optimized VMD to improve feature extraction effectiveness and reduce computational complexity. All components were normalized and then input into the TCN-BiGRU combined model for prediction. The prediction results of each component were superimposed and denormalized to obtain the final prediction value. Experimental results showed that compared with benchmark models， the proposed model had the best prediction accuracy， verifying its effectiveness， stability， and application potential.

Based on a multi-objective hierarchical optimization method， a hierarchical collaborative optimization model for the multi-energy complementary system in an offshore energy island was constructed， considering both system balance and economic efficiency of scheduling. To address the impact of the volatility and randomness of wind and solar energy on the scheduling of the offshore energy island system， the stochastic model predictive control （SMPC） method was adopted to optimize the scheduling of the offshore energy island. A scheduling method of adaptive variable step-size SMPC was proposed. In the rolling optimization link of SMPC， the real-time scheduling deviation degree was tracked through a deviation reference coefficient using the proposed method， and the rolling optimization step-size was adjusted accordingly. This addressed the problems of scheduling accuracy loss and local optimization in the rolling optimization phase of the traditional SMPC scheduling method， thereby balancing scheduling accuracy and globality. The simulation results show that this method can effectively improve the scheduling accuracy and shorten calculation time.

In recent years， the rapid development of digital technologies such as the Internet of Things， big data， and artificial intelligence has provided new methods for the operation optimization of integrated energy systems（IES）. A data-driven operation optimization method for IES was proposed. For an industrial park with a self-contained energy station in northern China， a deep learning long short-term memory neural network model was used for multi-load joint forecasting and photovoltaic power forecasting， providing accurate support for the operation optimization of the energy station. The main energy supply equipment was modeled under full operating conditions through data-driven machine learning algorithms. Taking energy efficiency， economic， and comprehensive benefit indicators as optimization objectives， the particle swarm optimization algorithm was applied to obtain typical daily operation optimization results. Under the condition of optimal energy efficiency indicator， the comprehensive energy utilization rate of the system reached 83.0%， and the operating cost was 64 802 yuan. Under the condition of optimal economic indicator， the system operating cost was as low as 64 590 yuan， and the comprehensive energy utilization rate reached 79.3%. Under the condition of optimal comprehensive benefit indicator， compared to the actual operation of the energy station， the comprehensive energy utilization rate increased by 7.5%， and the operating cost was reduced by 6 444 yuan. The results indicate that this operation optimization method has practical significance for guiding the operation optimization of IES.

The renewable energy-hydrogen-methanol integrated station（REHMIS） produces green hydrogen using electricity generated from renewable energy sources， and further synthesizes methanol from the green hydrogen and carbon dioxide， thereby achieving the substitution of green hydrogen for hydrogen produced from conventional fossil fuels. To simultaneously meet the methanol load demand of REHMIS and the multi-energy demand of its supporting buildings， a novel integrated energy system（IES） topology named REHMIS-IES was designed. To obtain an efficient operation strategy for REHMIS-IES， an execution framework based on the strictly constrained soft actor-critic（SC-SAC） algorithm was proposed. The established mathematical model was transformed into a Markov decision process， and a state constraint mechanism（SCM） was incorporated to prevent drastic fluctuations in the state of the energy storage system. In the execution stage of the SC-SAC algorithm， the trained Q-network and action constraints were transformed into a mixed-integer linear programming（MILP） model to ensure that scheduling decisions could comply with all operational constraints. The results from multi-scenario simulations showed that the proposed system could effectively reduce operating costs while meeting multi-energy demands. Compared with other deep reinforcement learning algorithms， the SC-SAC algorithm could lower the system energy imbalance by approximately 16.2% and reduce operating costs by at least 11.7%.

The heat load of integrated energy systems in parks is affected by multi-energy flows， and existing prediction models have insufficient feature extraction capabilities. To address this， a dual feature processing model for heat load prediction was proposed， integrating improved complete ensemble empirical mode decomposition with adaptive noise （ICEEMDAN） and multivariate phase space reconstruction. The heat load time series was decomposed using the ICEEMDAN method， and the components were reconstructed by calculating their sample entropy. These were then combined with input features such as air temperature to form multivariate time series datasets at different frequencies. The optimal delay time and embedding dimension of the series were determined using the C-C method， thereby obtaining the high-dimensional phase space of each dataset. The heat load components were predicted using a bidirectional long short-term memory neural network model with optimized parameters. The final heat load prediction value was obtained by summing the prediction results. The case study results showed that the proposed method achieved good prediction performance compared to other models.

With the large-scale integration of distributed photovoltaic （PV） systems into distribution networks， voltage fluctuations and out-of-limit issues have become increasingly prominent， and traditional single voltage control methods struggle to achieve rapid dynamic voltage control. Therefore， for distribution networks with high-penetration distributed PV integration， a robust control strategy based on the coordinated operation of distributed power source converters and static var compensators was proposed. Voltage control models for distributed power sources and static var compensators were established. Based on the control models of distributed power sources and static var compensators， a voltage sensitivity matrix was introduced to construct a unified voltage control model for the distribution network. By coordinating active and reactive voltage control， the active voltage regulation capability of the system was fully utilized to achieve rapid dynamic voltage control. Considering the system parameter uncertainties arising during distribution network operation， a robust control strategy was designed incorporating robust H_∞ performance constraints. Based on the IEEE 33-node system， case studies were conducted under scenarios of PV output fluctuations and sudden load changes. The results showed that the proposed strategy achieved rapid and stable voltage control and effectively suppressed voltage fluctuations caused by external disturbances， verifying the rapid response and effectiveness of the proposed control strategy.

In the context of the continuous advancement of big data technology and digital grid construction， the intelligent operation of the power grid faces both challenges and opportunities. To fully utilize the vast amount of data accumulated during power system operation， a new big data analysis and mining strategy was proposed in this study. Based on actual operational data from the power system， a big data analysis algorithm and visualization model were developed to conduct multidimensional statistical analysis from the perspectives of fault type and time scale， aiming to understand the operational status of distributed generation grids and the patterns of power quality faults， while further exploring their fault mechanisms. The research results showed that the proposed big data analysis and mining strategy effectively leveraged actual power system operational data， providing an intuitive understanding and analysis of power grid operational patterns and power quality fault characteristics， thereby offering strong technical support for the intelligent operation of distributed generation and high-quality power supply.

Aiming at the strong nonlinearity， non-stationarity， and multi-source coupling characteristics of island microgrid loads， a load prediction method was proposed， integrating robust empirical mode decomposition （REMD） based on evaluation factor reconstruction with detail-enhanced convolutional network （DECN） and bidirectional gated recurrent unit （BiGRU）. Multi-scale feature decoupling was achieved through REMD and evaluation factor reconstruction. A DECN-BiGRU hybrid architecture was constructed to fuse local differences and global dependency features， and multi-task learning was introduced to optimize the coupling relationships among components. Experiments showed that the model reduced the mean absolute percentage error by 68.78% compared with traditional methods and reduced the mean absolute error by 68.97% compared with deep learning models， thereby verifying the effectiveness of multi-modal feature fusion and bidirectional modeling. The research findings provide reference for power scheduling and energy storage configuration in island microgrids.