Aiming at the problem of diversified building energy system loads resulting in grid fluctuations and increasing load peak-valley differences, a grid peak-valley adjustment algorithm based on flexible loads and building multi-energy systems was proposed to achieve grid peak and valley adjustment with joint participation of multiple energy systems. To establish the objective function of power grid peak and valley adjustment for multi-energy system load and power optimization, power models of energy storage equipment and air conditioning systems were established. At the same time, to achieve power grid peak and valley adjustment, an improved particle swarm optimization algorithm was proposed to solve the objective function, according to the characteristics of power grid peak and valley adjustment. Simulation and experimental results show that the as-proposed method is superior to the grid peak and valley adjustment method, which uses only energy storage equipment and air conditioning system, and can significantly improve load smoothness and energy utilization.

Aiming at the disadvantages including high complexity and low accuracy of current security evaluation indexes for power system, a network security risk assessment method based on the Autoencoder method was proposed. By analyzing the vulnerability of system operation mechanism as well as by using the analytic hierarchy process and Delphi method, a preliminary assessment index system model of network security was established. By introducing the Autoencoder method, the complex index system model was simplified and optimized, and a new security assessment model for power system was formed. The simulation results show that the as-proposed model has higher execution efficiency and assessment accuracy, compared with the traditional security assessment model.

Aiming at the problem that the processing speed will be significantly slowed down when the Spark computing framework processes big data with rapidly increasing scale, and that it cannot meet the real-time processing requirements by power big data analysis, an analysis algorithm based on GPU and Spark computing framework for power big data was proposed. The parallel processing of GPU was combined with the Spark computing platform to improve the efficiency of power big data processing, and the performance of computing framework was maximized by building a queuing model. The simulation results show that the as-proposed algorithm has certain accuracy and effectiveness, can significantly improve the processing speed after adding GPU computation, and can meet the real-time requirements by large-scale data processing.

Aiming at the difficulties of power engineering cost analysis work, the large number of influencing factors and the uneven distribution of factor significance, a power engineering cost analysis system based on a KJ analysis method was proposed. The KJ analysis method was used to determine the main influencing factors of power engineering cost while reducing the number of unsignificant indicators. A combination weighting method was adopted to determine the weight of each influencing factor, and a ranking method for the approximation to ideal solution was used to reduce the uncertainty of indicators. In addition, a support vector machines was employed to analyze and evaluate the power engineering cost. The results of the case analysis for a 110 kV substation engineering show that the as-proposed method can improve the intelligence of the model, simplify the human operation, and obtain more reasonable evaluation results.

Aiming at the low accuracy of existing cost prediction methods for power transmission and transformation projects under complex working conditions, a cost prediction model based on BP neural network optimization algorithm for the power transmission and transformation projects was proposed. The input indexes for the data prediction of power transmission and transformation were determined by a factor analysis method. According to the traditional BP neural network model, the weight and threshold values of BP neural network were optimized by introducing a mind evolution algorithm. Using the as-proposed prediction model, the cost of power transmission and transformation projects in 2016 for a provincial power grid company was predicted. The results show that the prediction error is less than 10％, and the average error is less than 5％. Compared with the traditional BP neural network, the as-proposed prediction model has high prediction accuracy, and can be preferably applied to the cost estimation of power transmission and transformation projects.

In order to solve the problem that the existing abnormal data detection algorithms for electricity cost cannot identify list details and lists not inconsistent with construction details, a recognition algorithm based on rule matching for abnormal data of electricity cost was proposed. A K-means clustering algorithm was used to realize the preliminary classification of lists and the extraction of feature lists, and the feature words of feature lists were used as list category features. According to the rule base, the list details were segmented, and the list feature words were extracted. A polynomial Bayes algorithm was used to calculate the probability of a list in the current category. The experimental results show that the accuracy of as-proposed algorithm is about 10％ higher than that of traditional detection algorithms for abnormal data.

In order to solve the reconstruction problem of distribution network after failure, a practical and feasible multi-objective fault reconstruction scheme based on an improved universal gravitation algorithm was proposed. By investigating the fault reconstruction process of distribution network, the weight coefficients of multiple objectives were optimized, and a fault reconstruction model considering the network loss and the number of switching operations was proposed. By using reverse learning method, conjugate gradient method and Tent mapping, the classical universal gravitation algorithm was improved effectively. The results of software simulation and actual engineering tests show that the as-proposed universal gravitation algorithm has stronger practicability, lower network loss and fewer switching operations.

In order to determine the fatigue properties of narrow gap TIG welded joint of thick TC4 titanium alloy plate, the high cycle fatigue (HCF) tests at room temperature were carried out with the samples at three layers of welded joint, respectively. The morphologies of fatigue fracture surfaces and the microstructures near the fracture surface were observed and analyzed with scanning electron microscope and transmission electron microscope. The results show that the fatigue limits of three-layer welded joints are 420 MPa, 390 MPa and 400 MPa, respectively. For the first layer of welded joint, the fatigue crack initiation position is the sample surface, and the dislocation configuration is mainly dislocation network and dislocation array. For the second layer of welded joint, the fatigue crack initiation positions are both sample surface and welding defects, and the dislocation configuration is mainly dislocation tangles. For the third layer of welded joint, the fatigue crack initiation positions are sample surface and near-surface welding pores, and the dislocation configuration is mainly dislocation array and tangles.

In order to investigate the effect of heat treatment on the corrosion resistance of ZL114A aluminum alloy fabricated by wire-arc additive manufacturing (WAAM), the morphologies and chemical composition of second phases were analyzed by an optical microscope and a scanning electron microscope with energy dispersive spectrometer. The corrosion behavior was tested by an electrochemical test system. The results show that the microstructures of ZL114A aluminum alloy are mainly composed of α-(Al) matrix, eutectic Si and Mg₂Si phases after the solution and aging treatment. The morphologies of eutectic silicon phase gradually change to be spherical with the time increasing from 2 h to 14 h at 540 ℃ for solution treatment. The ZL114A alloy shows strong passivation ability after the solution treatment for 14 h. The amounts of eutectic silicon and Mg₂Si phase precipitated within grains gradually increase and the corrosion resistance of aluminum alloy decreases with the aging time changing from 4 h to 12 h.

In order to improve the quality factor value of MgTiO₃-(Ca_0.8Sr_0.2)TiO₃(MT-CST) microwave dielectric ceramics, the effects of Ce doping on the composition, crystal structure and microwave dielectric properties of MT-CST ceramics were investigated by SEM, XRD, Raman spectrometer and XPS analysis methods. The results reveal that the replacement of Mg²⁺ by high valence Ce³⁺ effectively inhibits the formation of oxygen vacancy and consequently decreases microwave dielectric loss, resulting in the great enhancement of quality factor value. The MCeT-CST ceramics have good dielectric properties with dielectric constant of 20.8, quality factor of 61 000 GHz and temperature coefficient of -4.99×10^-6/℃ at resonance frequency.

Aiming at the problem that the traditional passive suspension system cannot maintain good riding comfort under complex working conditions, a robust controller of electromagnetic active suspension was designed. The dynamic model of 7-DOF electromagnetic active suspension for whole vehicle was established, the robust control theory was used to weight the controlled output and input, and the H_∞robust controller was designed. The electromagnetic active suspension system was simulated and analyzed under the conditions of random road surface at the speed of 30 km/h and 60 km/h, as well as the sinusoidal pulse road surface at the speed of 20 km/h and 45 km/h, respectively. The results show that the main performance indexes of electromagnetic active suspension with robust control get improved compared with those of passive suspension. Under the random road surface, the vertical acceleration of vehicle body is improved by 28.03％, indicating that the designed robust controller is reasonable. The electromagnetic active suspension with robust control can effectively improve the riding comfort of the vehicle.

Aiming at the problem that conventional passive suspensions cannot utilize the energy of vibration, a linear motor type energy-harvesting suspension was proposed, and the energy-harvesting characteristics of the suspension were discussed. The quarter-vehicle dynamic model was established, and the suspension parameters influencing energy-harvesting characteristics were derived under sine excitation. Through taking the induction electromotive force as the index and based on the model for the linear motor type energy-harvesting suspension, the suspension parameters were simulated and analyzed with a control variate method. In addition, an experiment was conducted using by a quarter-vehicle suspension experimental platform. The results show that the sprung mass, unsprung mass and spring stiffness are negatively correlated with the induced electromotive force of the linear motor, and the tire stiffness is positively correlated with the induced electromotive force of the linear motor.

Aiming at the problem of low diagnosis rate for gear fault diagnosis when it is difficult to collect sufficient sample data, a method based on least square generative adversarial networks (LSGAN) and combined with long short-term memory networks (LSTM) was proposed. The original samples of gear were input into the LSGAN model, and the data were augmented, through the alternate training of the generator and the discriminator, to learn the sample data under different states. A LSTM diagnosis model was trained by the generated samples and the original samples to complete the fault diagnosis with small samples. The as-proposed method was validated by the experimental data on gears from the University of Connecticu. The results show that the diagnosis accuracy increases to 98.3％ compared with the traditional method. The visualization method shows the superiority of this diagnosis method and provides a reference for fault diagnosis under small sample conditions.

Aiming at the problem that the inclination defects of strain clamp in the transmission line cannot accurately perform the deblurring of acquired transmission line images, a quantitative detection method based on deep learning for the inclination defects of strain clamps in the transmission line was proposed. The obtained transmission line images were denoised by combining shock wave filter and bilateral filter, and the fuzzy kernel was estimated by a regularization method. At the same time, the guided filter was used to decompose the deblurred transmission line images to obtain both detail and basic layers, and the Gamma transform was performed respectively. The preprocessed transmission line images were input into the deep learning network, and the inclination defect features of strain clamps were extracted to finish the quantitative detection of the defects. The experimental results show that the accuracy of as-proposed method can be maintained above 90％, the recall rate, AP value and F₁ index are all relatively higher for satisfactory detection performance.

In order to reduce the Hamming loss and error rate of feature mining results, a dynamic feature mining algorithm of full supplier link based on decision tree was proposed. The data were numerically discretized and normalized, and the Hamming loss and error rate were reduced by reducing the difference of data range. The decision tree was established by using the processed data, and the accuracy of data classification was improved by pruning. The improved CHI value and RBF neural network were used to effectively mine the dynamic characteristics of the full link. The as-proposed algorithm reduces the Hamming loss and error rate of mining results, and improves the classification processing accuracy and recall rate. The as-proposed algorithm can effectively improve the mining effect of dynamic features of full supplier link.

In order to improve the effect of active disturbance rejection control(ADRC), an ADRC method based on improved krill swarm algorithm was proposed. Using a linear active disturbance rejection control technology, an active disturbance rejection controller of power system was designed. The expansion state observer was used to estimate the total disturbance of system, and the feedforward differential signal was added into the control quantity of system. The ADRC power system was controlled by the linear error feedback law of the controller, and the feedback proportional gain and feedback differential gain of the controller were optimized by the improved krill swarm algorithm to improve the ADRC effect. Experimental results show that the as-proposed method can effectively reduce the effective value of steady-state error and harmonic distortion rate of power system under different disturbances.

In order to solve the problem that the convergence of wireless sensor network coverage control is poorer and the adaptive ability is not strong, WSN network coverage question assumptions were proposed. According to the effective network coverage and the task node quantity, a WSN network coverage control model was built. In terms of particle swarm optimization (PSO) algorithm, an adaptive particle swarm optimization algorithm was proposed to obtain the best coverage, and the evolution and aggregation factors were added to the inertia weight coefficient for the improvement of PSO adaptability. Through the strategy of increasing collision rebound, the diversity of PSO was optimized, and the optimal coverage control of WSN network was realized. The results show that the as-proposed method has strong adaptive ability and good convergence of coverage control, and can reduce the energy consumption generated by the control of network node movement.

In order to investigate the bending performance variation of flange connected ductile iron pipes under acidic chlorine salt corrosion conditions, the bending test was performed after the flange-connected ductile iron pipes were treated by the alternation of wetting and drying in acidic NaCl solution, and the ultimate bearing capacity, yield bending moment, ultimate bending moment and strain of corroded specimens were compared and analyzed with the standard counterparts. The results show that the ratio of ultimate bending moment to yield bending moment of flange connected ductile iron pipes, which is finally about 1.2 to 1.4, becomes smaller with the increase of wetting and drying cycles. After 100 times of the alternation of wetting and drying, the ultimate bearing capacity of flange connected ductile iron pipes decreases by 35.13％. The flange connected ductile iron pipes treated by the alternation of wetting and drying in acidic chlorine salt solution is more prone to brittle damage at the connection between flange and pipe body.

In order to solve the problem of easy corrosion and difficult piling of the pile foundation of seamless beam bridge, taking a seamless quasi-continuous T-beam bridge as the research background, the overall model of the bridge and the simplified model of two-dimensional spring frame were established through finite element, and the finite element analysis of the mechanical performances of pile foundation was carried out. The horizontal resistance proportional coefficient m of each layer of foundation soil was converted into an equivalent m value, and the compressive stress and tensile stress of the normal section of the abutment RPC concrete (reactive powder concrete) cast-in-place piles under loading condition were compared and analyzed. The results show that m value and pile diameter have significant effects on the mechanical characteristics of the normal section of the abutment RPC concrete cast-in-place piles, and the tensile stress and compressive stress are the largest when the depth of pile section under the soil is 2.4D to 5.5D.