In traditional mobile edge computing systems, users face issues such as communication channel interruptions caused by dense obstacles and terrain structures, and under-utilization of idle system resources.These issues make it challenging to complete intensive computing tasks with low delay and power consumption.To address this, a UAV(Unmanned Aerial Vehicle)-assisted terminal pass-through collaborative mobile edge computing system is established. In this system, the computing user offloads part of the task to the idle user by establishing a direct connection link on the ground. The idle user uses their computing resources to assist with the calculation while offloading the remaining tasks to the UAV configured with a mobile edge computing server.A mathematical model of this new system is established, and a computational offloading strategy based on the deep deterministic policy gradient algorithm is proposed. This strategy optimizes the dual offloading rate and the maneuverability of the UAV to minimize the processing delay of computing tasks, under the constraints of UAV power and user movement range. Simulation tests in a simulated continuous state space environment show that the proposed offloading computing strategy optimization scheme can efficiently use resources in the collaborative network and effectively reduce task processing delay compared to other baseline algorithms.

To address the issues of motion distortion and error accumulation in SLAM(Simultaneous Localization and Mapping) systems during prolonged operations, a mapping method called IRIS-LOAM ( Lidar-Iris Based Lidar Odometry and Mapping in Real-Time ) is proposed, which leverages Lidar-Iris to construct global descriptors for loop closure detection. This algorithm has two major  innovations based on the LOAM algorithm.First, in the data processing stage, it integrates lidar data with IMU(Inertial Measurement Unit) data and uses the IMU data to correct the point cloud data. Second, in the mapping optimization stage, it employs an information matrix-based graph optimization algorithm and utilizes Lidar-Iris global descriptors for loop closure detection of key frames. And it preprocesses the input point cloud to improve the time efficiency of optimization. Comparing the improved algorithm with A _ LOAM through experiments, the results show that IRIS-LOAM achieves better mapping performance in various real-world scenarios, demonstrating its feasibility and practicality.

Big data usually comes from different data sources with diverse formats, structures, and qualities. Big data often contains a large number of redundant features, which can affect the accuracy of data classification during feature integration. To address these issues, a deep learning-based algorithm is proposed for feature integration classification in hospital big data. A feature extraction model is established based on deep learning to extract relevant features from the data. However, since the training process of the model introduces a significant amount of noise, the extracted features may contain irrelevant information, which can impact the results of feature integration classification. Therefore, a stacked sparse denoising autoencoder is employed to suppress irrelevant features. The best training parameters are determined using divergence functions and greedy algorithms, and a loss function is utilized to sparsify the irrelevant features in the feature space, resulting in practical data features.A feature integration classification model is constructed using an autoencoder network, and with the assistance of type-constrained functions and objective functions, the optimal integration centers for each class are obtained to achieve data feature integration classification. Experimental results demonstrate that the proposed method exhibits excellent classification performance, with macro-averaged values above 0. 95, and it also shows fast classification speed, indicating its effectiveness in classification.

In order to distinguish the effective component and noise component after VMD(Variational Mode Decomposition), and improve the denoising effect of VMD. A denoising algorithm (VMD-HD-VMD) combining VMD and HD(Hausdorff Distance) is proposed. Firstly, the original signal is decomposed into K IMF(Intrinsic Mode Functions) by VMD, the HD value of the probability density function of IMF component is calculated respectively, and the effective component and noise component are distinguished according to the HD value. Then the noise component is decomposed by VMD again, the effective component is selected by correlation coefficient, and reconstructed with the effective component decomposed for the first time. This method is applied to the denoising of pipeline leakage signal. The simulation experiment and pipeline leakage signal processing show that this method has better effect than EEMD(Ensemble Empirical Mode Decomposition), VMD and VMD combined wavelet denoising.

A proton magnetometer is a type of geomagnetic field measuring instrument based on the Larmor precession effect. However, a proton magnetometer with a single sensor is easily affected by geomagnetic diurnal variation and environmental interference. To improve the measurement accuracy of the proton magnetometer, sensor design and system performance evaluation are studied. Firstly, based on MAXWELL electromagnetic simulation software, simulation models of four kinds of sensors are established, including solenoid, cylindrical coil, ring coil, and "8" coil, and the directivity and anti-interference ability of the four coils are analyzed. Finally, it is determined that the "8" coil is used as the sensor to build a proton magnetic gradiometer. Secondly, the single and differential measurement results based on the fourth-order difference and mean square error algorithms are evaluated to analyze the performance of the magnetic gradiometer. The experimental results show that the initial signal-to-noise ratio of the "8"coil sensor can reach 40 / 1, the sensitivity of the single-channel mode can reach 0. 054 nT, and the sensitivity of the dual-channel differential mode can reach 0. 071 nT even under strong interference conditions, which is √2 times that of the single-channel mode.

In order to effectively improve the efficiency and performance of intelligent retrieval of book information resource services, a machine learning algorithm based intelligent retrieval method is proposed. Machine learning algorithms are used to classify the keyword frequency of books, providing the corresponding index for each class, calculating the weight of feature words corresponding to each document in the classification,and obtaining the score of feature words through repeated iterations. A contextual feature matrix is established,the corresponding score for each class is calculated, and the feature with the highest score is selected for classification processing. Based on the classification results of book information resources, grid computing technology is introduced to construct an intelligent retrieval model for book information resource services, and the model is used to achieve intelligent retrieval of book information resource services. The experimental results show that when using the proposed method for intelligent retrieval of book information resource services, the precision obtained is above 96. 0% , the recall rate remains above 90% , and the retrieval time under different datasets is around 450 ms, indicating that the proposed method has good performance in intelligent retrieval of book information resource services.

In order to effectively improve the visual effect and matching degree of text matching generated images, a text matching generated image model based on improved GAN( Generating Adversarial Networks) algorithm is proposed. Initial matching of text and images are unfolded through a mixed index tree. On the basis of GAN, they are improved to form an adversarial generation network based on cross attention mechanism encoding, and the improved GAN is used to establish a text matching image generation model. The cross attention encoder in the bidirectional LSTM( Long Short-Term Memory) network optimization model is used to translate and align text and visual information, obtaining cross modal mapping relationships between text and images, completing fine matching between text and images, and ultimately generating images that meet the requirements of the text. The experimental results show that the proposed model can generate images with higher quality that match image details with text.

With advancements in modern science and technology across structural design, materials, and sensor manufacturing processes, pipeline leak detection sensors are becoming increasingly miniaturized and intelligent, and technologies for measuring physical changes caused by pipeline leaks continue to mature. For the convenience of technology selection and optimzation in pipeline leak detection, a systematic review of widely used piezoelectric, optical fiber, and laser sensors in pipeline leak detection is provided, with a focus on analyzing their characteristics and differences in materials, structures, and working principles. It further explores their performance in practical applications and the current state of research both domestically and internationally, offering theoretical support and technical references for the selection, optimization, and future development of pipeline leak detection technologies.

As the most widely adopted backbone network in classification, object detection and instance segmentation tasks, the representation capability of ResNet ( Residual Neural Network) has gained extensive recognition. However, there are still certain limitations that hinder the representation ability of ResNet, including feature redundancy and inadequate effective receptive field. To address these problems, a feature fusion block is proposed, which can fuse features of different scales to construct multi-scale features with richer information and improve channel utilization, when the model channel is increased. The block employs a small number of large kernel convolutions, which is benefit to the expansion of the effective receptive field of the model and the trade-off between performance and computational efficiency. And a lightweight downsampling block and a shuffle compression block are also proposed, which can effectively reduce the parameters of the model and make the entire method more efficient. The feature fusion block, downsampling block and shuffling compression block are introduced to the ResNet can build a FFNet(Feature Fusion Network), which will have faster convergence speed and a larger effective receptive field and better performance. Extensive experimental results on CIFAR (Canadian
Institute for Advanced Research ), ImageNet and COCO ( Microsoft Common Objects in Context ) datasets demonstrate that the feature fusion network can bring significant performance improvements in classification, object detection and instance segmentation tasks while only adding a small number of parameters and FLOPs(Floating Point Operations).

To address the challenge of efficiently mining spatiotemporal information for traffic prediction, a novel vehicle travel time prediction method is proposed based on bidirectional multi-attention spatiotemporal graph convolution. To extract the spatial dependencies within the road network, a traffic transfer matrix is constructed using a Markov chain approach, which captures the bidirectional traffic flow transfer relationships. Graph convolution is employed to learn the spatial dependencies within the graph network. Subsequently, an attention mechanism is utilized to capture both local and global temporal features within the traffic flow map. Finally, a MLP ( Multi-Layer Perceptron) is used to forecast travel times, producing the final prediction results. The Xuancheng road network traffic data is selected for model validation. The results demonstrate that the proposed model reduces the RMSE (Root Mean Square Error) by 7. 6% , 3. 7% , and 9% , respectively, compared to baseline models such as STGCN( Spatio-Temporal Graph Convolutional Networks), ASTGCN(Attention Based Spatial-Temporal Graph Convolutional Networks), and A3T-GCN ( Attention Temporal Graph Convolutional Network). This significant reduction in RMSE indicates that this model substantially improves prediction accuracy, highlighting its effectiveness in capturing and utilizing spatiotemporal information for more precise traffic predictions.

Because there are many factors affecting the network link state, it is difficult to accurately judge the network link state and the location where congestion is about to occur, resulting in the decline of the quality for network link congestion control. A priority based mathematical modeling method for network link congestion automatic control is proposed. A heterogeneous network model combined with the current occupation and change of network interface queue buffer space is built to detect the network link congestion, and judge the network link status and the location where congestion is about to occur. Through the data request sending mechanism, the
detection results are fed back to other network nodes in time. The high priority burst packets in the network nodes are set to have absolute priority. The burst packets of the network link are divided and deflected to realize the automatic control of network link congestion. The experimental results show that the model can effectively detect the current network link congestion and control the congestion degree of the network. It has better network communication data transmission effect, blocking rate, better throughput and higher control quality.

There are many types of information features in the supply chain privacy network, and it is difficult to determine data ciphertext, which leads to high encryption difficulty. Therefore, a chaotic sequence based information encryption algorithm is proposed. Considering the risk of privacy information being leaked or tampered with, the risk variation is used as a reference feature for encryption algorithms. The risk index is obtained by solving the same, different, and opposite risk variation characteristics of privacy information through ternary and quinary functions. A chaotic sequence is established based on the risk index of the information to be
encrypted, a key sequence is generated through the chaotic sequence, the dimensions corresponding to different types of information are calculated, and finally the information encryption is achieved through chaotic mapping between the information and the key sequence. Experimental data shows that the proposed algorithm has high accuracy in encrypting supply chain privacy network information, low packet loss rate, and can effectively improve the phenomenon of privacy information leakage and loss during transmission.

In order to ensure the normal operation of the ventilator, an intelligent monitoring method for the operating status of the ventilator based on the HHT(Hilbert-Huang Transform) algorithm is proposed. Firstly,wavelet neural network is used to denoise the running signal of the ventilator; Secondly, combined with the HHT algorithm, the denoised ventilator operation signal is decomposed by EMD(Empirical Mode Decomposition), and the decomposed IMF( Intrinsic Mode Functions) component is transformed by Hilbert spectrum to obtain the signal spectrum as the signal feature. Finally, the obtained signal spectrum is placed in the MLP neural network
classifier, and the backpropagation algorithm is used to train the MLP neural network to achieve recognition of the operating status of the ventilator. The experimental results show that the proposed method has a good denoising effect, and the monitored results are consistent with the actual spectrum. At the same time, the sensitivity of monitoring is above 96% , and the accuracy of operating status recognition is above 95% . This indicates that the proposed method can effectively monitor the operating status of the ventilator and has good monitoring performance.

For the classification of networked data, most relational network classifiers are based on the homophily hypothesis, and the simplified processing based on the first-order Markov assumption has certain limitations. The local graph ranking algorithm ( activation spreading) is introduced into the network Bayes classifier instead of the original direct neighborhood acquisition method. The neighborhood range of nodes to be classified is appropriately expanded by setting the initial energy and the minimum energy threshold, increasing the homophily of nodes. Combined to the collective inference method of relaxation labeling, the classification
accuracy of network data is improved to a certain extent. Compared to 4 network classifiers, the experimental results show that the classification performance of the proposed method on 6 networked datasets is improved in different degrees.

Air combat is an intense, complex, and continuous process, filled with many influencing factors,complex interactions, and uncertainties. Aiming at the problem of low accuracy of UAV ( Unmanned Aerial Vehicle) cluster air combat effectiveness assessment, a method based on System Dynamics is proposed. The interactions between various factors is analyzed in the UAV air combat system during a reconnaissance and strike mission carried out by the red formation. A causal relationship model and a stock flow model are established.The evaluation indicators of combat effectiveness are constructed based on three aspects: combat effect, combat
efficiency, and combat cost. These indicators include mission completion, combat efficiency, and combat loss rate. Then, the influencing factors of the air combat system and the combat effectiveness under different equipment schemes are studied. finally the feasibility and effectiveness of the proposed method are verified through simulation, which solves the complexity and uncertainty problems arising in the assessment process.

When solving high-dimensional equations, the CS (Cuckoo Search) has the drawback of falling into local optima. To address this deficiency, an improved hybrid cuckoo search is proposed. Firstly, the population is initialized using chaotic mapping and reversed learning mechanisms. Then the search mechanisms of TLBO (Teaching Learning Based Optimization) and CS are performed alternately. Finally, the discovery probability and embeds DE (Differential Evolution) are dynamically adjusted to comprehensively improve the algorithm's performance. The comparative results of simulation experiments with 6 benchmark functions and 1 optimized grating coupler design show that this algorithm is better for solving high-dimensional equations and effectively avoids the CS algorithm getting stuck in local optima.

In order to improve the safety and efficiency of tunnel operation, a multi-sensor based tunnel operation risk perception method is proposed. Firstly, an object detection method based on YOLOv7(You Only Look Once v7) is used to effectively detect information such as traffic flow and speed from video sensors, and a deep temporal convolutional network algorithm is used to dynamically evaluate tunnel operation risks. Video sensors are integrated with multiple sensing devices such as smoke and temperature sensors to form a fire risk monitoring system, and multiple sensing devices are integrated to form a fire monitoring system. By numerically modeling
the fire monitoring status under different combustion conditions inside the tunnel, the distribution and development patterns of temperature, toxic gas concentration, and smoke inside the tunnel are analyzed, providing a risk assessment of tunnel fires. This method is applied in some tunnels of Shandong Jiwei Expressway to ensure the safety of the tunnels.

Deep learning methods have improved the efficiency and accuracy of fault identification, but current research often relies on extracting fault features from single attributes such as seismic amplitude, which leads to issues like poor fault continuity and missed detections. These problems limit the exploration and development of oil and gas reservoirs in complex areas. An intelligent fault identification method based on deep learning technology is proposed, which adopts a multi-level fusion strategy to construct a dual-attribute parallel fusion network PE-Net(Parallel Elements Network). Firstly, the ant body attributes and amplitude attributes are input
into the ant body feature extraction network and the amplitude feature extraction network respectively, capturing the fault features of different angles from both paths using the AIFM ( Attribute Intensive Feature Module). Secondly, two attribute feature modules are used to integrate cross-layer features of the output of each branch, mining multi-scale information and mitigating scale changes. Finally, the FFM(Feature Fusion Module) is used to integrate the two parallel branches, reducing the limitation of a single attribute. Synthetic data experiments demonstrate that the PE-Net model achieves an accuracy of 97. 95% , with a 1. 33% improvement compared to the U-Net model. The fault identification results on the Kerry3D dataset and ablation experiments confirm that the proposed method is capable of capturing more contextual fault features, reducing missed and false detections,thereby improving the accuracy of complex fault identification and enhancing the detection of small faults.

As the environment of geophysical exploration becomes more complex, and it is limited by acquisition and processing technology which results in low resolution and signal-to-noise ratio of post-stack seismic data.Therefore, how to enhance its resolution while realizing noise attenuation is a non-negligible problem. A design called DESRGAN(Depth-Enhanced Super-Resolution Generative Adversarial Network) is proposed, which is intended to be applied to the task of super-resolution reconstruction of seismic data. DESRGAN uses a LRDB(Lightweight Residual Dense Block) as the base unit to improve the efficiency and stability of the training
process, passes through channel attention in the deep feature extraction phase to increase the focus on important features and performs an up-sampling operation using pixel reorganization instead of interpolation to take into account the spatial relationship between pixels. Experimental results on synthetic and field data show that the network can reconstruct the synthetic data as the test set and it is well generalized to the field data. Compared with classic GAN(Generative Adversarial Network) and CNN(Convolutional Neural Network), the reconstructed results are visually clearer, and have higher peak signal-to-noise ratio and structural similarity in quantitative analysis.

 Due to the Irreversible process of scrambling and extraction process, the hidden information of digital media video image information can not be completely recovered in the extraction process, resulting in information loss or error, and reducing the effectiveness of the hiding algorithm. To solve this problem, a digital media video image information hiding algorithm based on reference frames is proposed. Firstly, the NLEMD(Neighborhood Limited Empirical Mode Decomposition) algorithm is used to enhance digital media video images and improve video image quality. Secondly, the Arnold transform scrambling method is used to perform scrambling transformation on the enhanced image, completing the preprocessing of information hiding. Finally, the scrambled digital media video image information is hidden through an information hiding algorithm based on reference frames. The experimental results show that the proposed method can improve the peak signal-to-noise ratio of digital media video images, reduce the time required for embedding and extracting hidden information,and achieve high accuracy in information extraction.

Existing Chinese named entity recognition algorithms inadequately consider the data features of entity recognition tasks, leading to imbalance in the categories of Chinese sample data, excessive noise in the training data, and significant differences in the distribution of generated data. An improved Chinese named entity recognition model based on BERT-BiLSTM-CRF ( Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory-Conditional Random Field) is proposed. The first improvement involves combining the P-Tuning v2 technology with BERT-BiLSTM-CRF to accurately extract data features. And three
loss functions, including Focal Loss, Label Smoothing, and KL Loss(Kullback-Leibler divergence loss), are utilized as regularization terms in the loss calculation to address the problems. The improved model achieves F1 scores of 71. 13% ,96. 31% , and 95. 90% on the Weibo, Resume, and MSRA( Microsoft Research Asia)datasets, respectively. The results validate that the proposed algorithm outperforms previous research achievements in terms of performance and is easy to combine and extend with other neural networks for various downstream tasks.

In order to explore the analysis of research points and theme evolution of science and technology resources in China. Bibliometrics and knowledge mapping analysis tools are used to statistically and visually analyse the research papers on S&T resources in China in the past 10 years. The results show that China's technological resources resources are in the growth period, with China Institute of Science and Technology Information, National Science and Technology Basic Condition Platform Centre and Guilin University of Science
and Technology as the main issuing institutions, Research on Science and Technology Management, China Science and Technology Resource Guide, and Science and Technology Progress and Countermeasures as the main journals carrying the papers, and the hotspots of research are the allocation of S&T resources and the influencing factors, the sharing and integration of S&T resources, and the construction of the resource-sharing platform and the The research hotspots are S&T resource allocation and influencing factors, S&T resource sharing and integration, resource sharing platform and service platform construction, collaborative innovation is the research hotspot after 2014, and S&T resource structure, innovation chain and S&T resource pool are the cutting-edge hotspots in the recent three years.

Aiming at the limitations of existing named entity recognition methods in identifying entities involving multiple elements and nested entities in oil and gas exploration texts, a novel approach is proposed. This approach integrates multiple features using a BERT-CNN-BiGRU-Attention-CRF(Bidirectional Encoder Representations from Transformers-Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Attention-Conditional Random Field) architecture for named entity recognition. The model leverages BERT's semantic extraction capability to obtain character Vectors with global features for the entire sentence. Additionally, it utilizes CNN's ability to capture local features, overcoming limitations of BERT character Vectors, and obtains character-level Vectors for words. By incorporating a custom oil and gas exploration domain dictionary and employing a bidirectional maximum matching method, dictionary feature Vectors are obtained. These three types of Vectors are concatenated and used as input for the BiGRU-Attention-CRF model. Experimental results on a self-constructed small-scale oil and gas exploration dataset demonstrate an F₁ score of 91.10% . Compared to other mainstream NER ( Named Entity Recognition) methods, this model exhibits superior recognition performance. Furthermore, it provides valuable assistance in constructing knowledge graphs for the oil and gas exploration domain.

AI(Artificial Intelligence) plays an important role in medical imaging education, driving innovation in teaching methods and medical education. With the continuous development of AI technology, especially breakthroughs in deep learning, image recognition, and natural language processing, AI is gradually demonstrating its unique advantages in the field of medical imaging education. AI helps students and healthcare professionals quickly identify disease features, and provides automated image analysis results, allowing learners to intuitively understand the imaging manifestations of different diseases. It enhances the interactivity and practicality of learning. AI can offer personalized learning paths recommending relevant educational content or exercises based on the student’s progress and understanding, ensuring that learners receive tailored educational services. The efficiency and accuracy of AI assist students in better comprehending complex medical imaging content improving learning outcomes. However, AI in medical imaging education also faces certain challenges.As technology continues to advance, AI will play a more significant role in medical imaging education. Future educational systems are likely to become more intelligent, integrating technologies such as virtual reality and augmented reality to provide students with a more immersive learning experience.

To address the efficiency and quality issues in current molecular diagnosis talent cultivation, the application status and future development trends of AI(Artificial Intelligence) technology in molecular diagnosis talent cultivation is explored. The research content covers the current application of AI technology in molecular diagnostics, its advantages and challenges, and focuses on analyzing how AI can enhance the efficiency and quality of talent cultivation through automated experimental processes, precise data analysis, and
interdisciplinary knowledge integration. The study summarizes practical experiences from domestic and international universities in integrating AI with molecular diagnostic talent cultivation and outlines future development trends, including the integration of VR ( Virtual Reality ) and AR ( Augmented Reality )technologies, the precision of intelligent diagnostic systems, and the intelligence of personalized learning platforms. The conclusion of the study indicates that AI technology holds great potential in the cultivation of
molecular diagnostic talents, significantly enhancing their comprehensive competitiveness and promoting the further development of molecular diagnostic technologies to provide robust talent support for precision medicine. However, the application of AI technology still faces multiple challenges, including the integration of interdisciplinary knowledge, data quality, and ethical and privacy issues, which need to be addressed through the joint efforts of educational institutions, industries, and governments.

The specific high complexity and nonlinear characteristics of large-scale road network traffic flow in a short period of time affect the accuracy of short-term traffic flow prediction. A short-term prediction method for large-scale road network traffic flow is studied based on improved neural network algorithms. Large-scale road network functions are constructed, road network functions are optimized by treating road sections as the core of the network and treating road nodes as corresponding connecting elements. Based on the optimized road network function, traffic flow features are extracted by combining the K-means algorithm with the EM ( Expectation-Maximization) algorithm. By combining genetic algorithm with Elman neural network algorithm, a short-term prediction of the traffic flow of the road network is carried out, and relevant prediction results are obtained.Experimental results have shown that the improved method's single point average speed prediction results are closer to the actual values, and the short-term prediction error of large-scale road network traffic flow is lower,resulting in higher reliability of the prediction results.

The scale of multimedia digital image data is enormous, including ordinary RGB(Red Green Blue)images and various types of data such as depth maps, texture information, and normal maps. The difficulty of depth estimation leads to low accuracy in 3D scene reconstruction. To effectively address this issue, an interactive 3D virtual scene algorithm for multimedia digital images is proposed. Corners from multimedia digital images are extracted and they are used as initial values to perform a checkerboard edge search to determine the true corners. All corners are used as feature points to perform camera calibration and obtain the corresponding
camera pose for each multimedia digital image. By using the improved PatchMatchNet to perform depth estimation on the reference image, the output depth map is obtained through multiple iterations. By using the method of reprojection to filter the outer points of the depth map and projecting it into the world coordinate system, an interactive 3D virtual scene is finally obtained. The experimental results show that the proposed algorithm can obtain high-precision interactive 3D virtual scene reconstruction results, and the reconstruction time is less than 50 ms.

To ensure stable transmission of big data, a device big data transmission load optimization algorithm based on 5G network is proposed. The factors that affect the performance of big data transmission are analyzed,including data latency, average stream bandwidth utilization, and throughput. Morphological filtering algorithms are used to perform low-pass filtering on big data, eliminating noise in the data and reducing data transmission delay. Dynamically big data transmission channels are selected to avoid data congestion in the network and improve network throughput. On the basis of information transmission matrix mapping, data transmission accuracy is improved, and a capacity expansion mechanism is designed to improve network bandwidth utilization
and complete load optimization. The experimental results show that after optimization using the proposed algorithm, the bandwidth utilization rate is improved, and the network energy consumption and data transmission delay are reduced.

In order to better utilize solar energy, it is very important to accurately predict photovoltaic power generation. To improve the accuracy of photovoltaic power prediction,a photovoltaic power prediction method based on the combination of factor-related complementary ensemble empirical mode decomposition and optimized long short-term memory network is proposed. Firstly, the CEEMD(Complementary Ensemble Empirical Mode Decomposition) algorithm is used to decompose the photovoltaic power sequence, and the Pearson correlation coefficient matrix of the decomposed power components and environmental factors is established. Three key factors are selected as the input of the subsequent prediction for each decomposed power component.