To address issues such as diverse N-k cascading fault scenarios， complex fault propagation paths， and difficulties in determining protection strategy implementation targets， a critical section identification and protection configuration model that integrates XGBoost with Bayesian hyperparameter optimization is proposed in power systems. By constructing a high-risk N-k fault set and randomly simulating cascading faults under load rates ranging from 0.1~10.0， a cascading fault dataset was developed using line load rates as inputs and residual load as the target. The Bayesian optimization algorithm was used to fine-tune the hyperparameters of the XGBoost model， selecting the optimal parameter combination. The protection resource allocation strategies for high-risk N-k fault scenarios were then identified. Simulation results on the IEEE 39-bus system showed that for 88% of high-risk N-k fault scenarios， adjusting the power flow carrying capacity of three critical section lines enabled the system's residual load to remain above 80%.

With the increase of new energy sources and converter feed-ins， the equivalent inertia of a power system continuously decreases， leading to instable power systems with high-proportion new energy and massive electronic devices accessed. Critical inertia，the lower limit of the inertia， represents safety boundary of grid frequency，and it is a key parameter in judging the safety of the power grid. However， the solving process is complex since the parameter is nonlinear. In view of this， a linearized calculation method for critical inertia of the energy system with renewable energy is proposed. By introducing the critical inertias corresponding to the two boundary constraints， the lowest frequency and the maximum frequency variation rate， the equivalent rotor motion equation of the grid is established，and the primary frequency response process of a synchronous generator set is linearized， realizing the linearized analytical solution for the critical inertia. Finally， the accuracy of the calculation results is verified through PSASP software simulation executed on an actual power grid model.

Timely early warning of security situation in cyber-physical systems （CPS） of urban power grids is critical for ensuring safe and stable operation. To address the early-warning challenges for the operational status of CPS under multiple disturbances， a security situation early warning method based on kernel extreme learning machine （KELM） was proposed. A coupling model of the physical and information layers of the power grid was established by integrating cellular automata theory， and the mechanism of cross-space risk propagation was analyzed； An ensemble KELM early warning model was developed， in which multidimensional data were deeply integrated through radial basis function kernel mapping， and prediction accuracy was enhanced by the ensemble structure； An early warning indicator system was established， and indicator weights were dynamically allocated using the entropy weight method to classify early warning levels of security situation. Simulation experiments based on the IEEE 33-bus distribution network demonstrated that， under distributed generation integration scenarios， the proposed method achieved a 12.49% reduction in mean squared error of voltage fluctuation prediction compared to traditional extreme learning machine methods， verifying the efficiency and robustness of the model.

Due to the intermittency and instability of solar radiation， photovoltaic（PV） power generation shows high randomness and fluctuation， posing challenges to the stable operation of power grids. To improve prediction accuracy， the complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN） was applied to decompose PV power data into intrinsic mode functions（IMFs） with different frequencies. These IMFs were clustered using K-means based on sample entropy， categorizing into high-， medium-， and low-frequency components. The high-frequency components were further decomposed using variational mode decomposition（VMD） for refined analysis. A hybrid deep learning prediction model was established by integrating convolutional neural network（CNN）， bidirectional long short-term memory （BiLSTM） network， and attention mechanism. The osprey optimization algorithm（OOA） was employed to optimize the model's hyperparameters. The experimental results showed that the proposed hybrid prediction model based on modal secondary decomposition and OOA-CNN-BiLSTM-Attention achieved a root mean square error of 4.11 kW， a mean absolute error of 2.88 kW， a mean absolute percentage error of 3.08%， and a coefficient of determination of 98.89%， outperforming other models. It is demonstrated that the proposed method effectively captures the multi-scale features of PV power generation with strong generalization ability and application potential.

To enhance the prediction accuracy of wind turbine output power under diverse meteorological conditions， a power regression prediction method based on complete ensemble empirical mode decomposition with adaptive noise （CEEMDAN）， convolutional neural network（CNN）， and long short-term memory（LSTM） network was proposed. The CEEMDAN algorithm was employed to decompose the original wind power data into intrinsic mode functions（IMFs） and a residual（RES）. Five meteorological factors， including wind speed， were incorporated， and CNN was applied to extract features. LSTM networks were used to perform regression prediction for each subsequence. The prediction results were then superimposed and reconstructed to obtain the final predicted values. Prediction accuracy was evaluated using mean absolute error and root mean square error. Simulation results indicated that the CEEMDAN-CNN-LSTM model significantly outperformed the random forest-LSTM（RF-LSTM） and support vector machine-LSTM（SVM-LSTM） models in prediction accuracy， with notably improved performance and generalization capability under complex meteorological conditions and extreme weather events.

Accurate prediction of photovoltaic power generation under extreme weather events is crucial for ensuring energy supply and grid stability. However， the suddenness of such weather events leads to scarce historical data from photovoltaic power stations， making it difficult to effectively predict photovoltaic power under extreme weather conditions. To address this issue， a prediction method based on Time-series Generative Adversarial Networks（TimeGAN） was proposed to augment limited historical data. The method captured the complex temporal dependencies between photovoltaic power and weather conditions. Based on the limited historical data from photovoltaic power stations， the TimeGAN model generated realistic time-series data to simulate the occurrence of extreme weather events， and subsequently conducted photovoltaic power prediction. The experimental results showed that compared to traditional GAN for small sample augmentation， the TimeGAN-augmented prediction results demonstrated better fitting performance. After 25% data augmentation， the Mean Absolute Error（MAE） decreased by 1.14 MW， and the Root Mean Square Error（RMSE） decreased by 1.09 MW. After 50% data augmentation， the MAE decreased by 1.08 MW， and the RMSE decreased by 0.99 MW. These results indicated significant improvements in prediction accuracy.

To address the challenges of insufficient multi-source heterogeneous data fusion and uncertainty quantification in load forecasting for integrated energy systems， a hybrid CNN-BiLSTM-RF-KDE model was proposed. The model utilized convolutional neural network（CNN） to extract local features of load data， bidirectional long short-term memory （BiLSTM） to capture bidirectional temporal dependencies， random forest（RF） to handle high-dimensional nonlinear relationships， and kernel density estimation（KDE） to quantify prediction uncertainty. Additionally， an electricity-heat-gas multi-energy flow coupling model was established to analyze the influence of different carbon price intervals on scheduling strategies. Case studies demonstrated that the coefficient of determination（R²） for electrical and heating load forecasting reached 0.93 and 0.96 on the training set， and 0.79 and 0.84 on the test set， respectively. The predicted power generation and heat output of each device closely aligned with the mean trend， indicating that the model provided more accurate load predictions. Based on this data， more reliable analysis and scheduling of integrated energy systems could be achieved.

To address the limitations of traditional differential evolution （DE） algorithms—susceptibility to local optima and weak physical interpretability—in the optimal configuration of wind-solar-storage systems in multi-park microgrids， an optimization framework integrating an improved DE algorithm with physical mechanisms was proposed. A wind-solar-storage configuration model was established with the objective of minimizing daily power supply costs， incorporating constraints on energy storage charging/discharging efficiency and state of charge. A triple-adaptive improved DE algorithm was designed： a dual-phase linear decay mechanism was used to adjust the scaling factor and crossover probability， an elite historical experience reuse strategy was integrated to enhance convergence speed， and a dual-mode oscillatory disturbance was introduced to increase population diversity. From the physical essence of source-load matching， the intrinsic patterns between energy storage configuration and wind-solar load curves were analyzed. Case studies showed that： the improved DE algorithm outperformed traditional DE， particle swarm optimization， and genetic algorithms， reducing the joint operating costs to 15 424.06 yuan； the joint operation reduced the power supply cost by 6.11% compared to the sum of independent operations and saved 30.77% of total energy storage power and 50.00% of capacity， with wind and solar curtailment reduced to zero； energy storage configuration followed a universal pattern that the power rating was determined by the maximum single-interval curtailment， and the capacity was determined by the maximum consecutive curtailment. Based on this， the physical estimation scheme for Park B reduced the coet to 5 065.43 yuan， which was lower than the result obtained by the optimization algorithm （5 066.22 yuan）. By combining algorithm improvement with the exploration of physical patterns， a high-precision and strongly interpretable solution for the optimal configuration of wind-solar-storage systems is provided.

To investigate the Stackelberg game when electric vehicles （EV） participate in virtual power plant （VPP） scheduling and to analyze the effect of conditional value at risk （CVaR） on VPP profits， a bilevel Stackelberg game scheduling model for EV-integrated VPP considering CVaR was established. A bilevel Stackelberg game model between the VPP and EV was established， in which the VPP acted as the upper-level leader， setting electricity prices and guiding the lower-level EV to carry out orderly charging. Then， CVaR theory was introduced into the upper level to measure the risks brought by renewable energy such as wind and solar power. An upper-level objective function was established to maximize VPP profits and minimize CVaR risks， and a lower-level objective function was set to minimize the charging costs of EV. The proposed Stackelberg game scheduling model was solved using MATLAB + CPLEX solver. The results showed that different participation ratios of EV in the VPP had different effects on the total VPP profits. The highest VPP profits were achieved when EV types were concentrated. The established model can effectively guide the VPP to avoid risks while minimizing the charging costs of EV.