The result of the present electricity price prediction as a key signal in the power market,for the normal operation of the power system,plays an important role.In this paper,a prediction model of the present electricity price is proposed based on the self-attention mechanism with the long-nosed raccoon optimization algorithm of the convolutional neural network and the bi-directional gated recurrent unit network.The model fully considers many factors such as the boundary conditions of the power market and the external environment that affect the price of electricity,and firstly uses the Pearson correlation coefficient method to correlate the disclosure data of the power market in Shandong Province,and comes up with the key factors that affect the price of electricity.Then the data are input into a CNN-BiGRU model based on self-attention mechanism and long-nosed raccoon optimization algorithm for training.The experimental results show that the three evaluation indexes of the model,Mean Absolute Error (δ_MAE),Mean Absolute Percentage Error (δ_MAPE),and Coefficient of Determination (R-square,R²),are 10.481, 3.23%, respectively, 0.954,the three indicators are obviously better than other models,with higher prediction accuracy and stability,fully verifying the feasibility of the model in the prediction of the present electricity price.

The correct disposal of monitoring and alarm information in substation is very important to ensure the safe operation of substation and equipment maintenance.The premise of correct disposal is to extract the power entities in the monitoring alarm information quickly and accurately.Because of the existing power entity extraction methods,it is difficult to meet the dual requirements of accuracy and speed in practical application.In order to solve the above problems,a semi-supervised-ERNIE-GP power entity extraction method based on enhanced representation through knowledge integration (ERNIE) and Global Pointer (GP) and using semi-supervised learning strategy is proposed.This method is based on ERNIE-GP to improve the accuracy and speed of power entity extraction,and introduces semi-supervised learning idea to mine entity extraction knowledge from unlabeled data.In order to verify the effectiveness of the proposed method,the monitoring and alarm information of substation is used to construct a data set and conduct a series of experiments.Comparative experiments show that,compared with the better baseline model BERT-Bi-LSTM-CRF,the Semi-Supervised-ERNIE-GP adopted in this thesis improves the accuracy,recall and F1 score by 4.90%,2.50% and 3.71% respectively.Through curve analysis,the superiority of the extraction speed of this method in large-scale data application scenarios is further confirmed.

With the continuous development of medical imaging technology,chest CT images play a crucial role in the early diagnosis and treatment of lung diseases.Computer-aided detection systems can provide valuable references for clinicians,thereby reducing diagnostic errors caused by human factors.To address the challenge of varying feature channel importance in lung cancer segmentation from chest CT images,TransUnet-SE is proposed,this net is an enhanced Transformer-based U-Net architecture incorporating residual-aware mechanisms for pulmonary lesion segmentation.The SENet attention mechanism is embedded into the decoder’s upsampling process to accurately mitigate multi-channel feature differences through a three-step process of ‘squeeze,excite,and scale’.To validate the model’s generalizability,EMPIRICAL TEST was firstly conducted on the public Synapse multi-organ CT dataset,followed by fine-tuning and evaluation on lung cancer-specific CT images from the Lung-PET-CT-Dx dataset.Comparative experiments with state-of-the-art models demonstrate that our method achieves a Dice Similarity Coefficient of 86.05%.Furthermore,a user-friendly lung cancer segmentation assistant system was developed using PyQt5 for clinical implementation.This system invokes the weight parameters of the TransUnet-SE model to implement the segmentation function,thereby providing support for clinical diagnosis.

To address the low detection accuracy caused by occluded pest targets and camouflage effects (where pest body colors blend with the environment) in crop pest detection,this study proposes an RT-DETR-based algorithm named RT-DETR-SDIC.First,the original backbone network’s early stages (S2,S3) are augmented with DBRB.By integrating multi-branch topological structures and heterogeneous paths of varying scales and complexities,the DBRB enriches the feature space.The later stages (S4,S5) of the backbone network are enhanced with IRMB_CGA.This module mitigates the lack of direct long-range semantic interaction in the original architecture while improving discrimination capability for environmental features.Second,in the feature fusion network,a parameter-free attention mechanism— SPA—is introduced to capture fine-grained spatial information.Additionally,CGFM is proposed for the feature fusion layer,orchestrating multi-scale feature integration.Experimental results demonstrate that RT-DETR-SDIC achieves a 19.6% reduction in parameters,a 9.9% decrease in computational load,a 6.2% improvement in P_mA,0.5 (average precision at IoU=50%),and a 2.6% improvement in P_mA,50:95 (mean average precision across IoU thresholds from 50% to 95%).

To enhance vehicle ride comfort,an active suspension system based on the inverse piezoelectric effect with piezoelectric stacks is proposed.According to Lagrange’s theorem,a six-degree-of-freedom active suspension dynamics model is established.Six performance metrics including vertical cabin acceleration are selected to formulate a fitness function,and an LQR (Linear Quadratic Regulator) controller is designed by integrating a genetic algorithm.Numerical simulations demonstrate that the piezoelectric effect-based active suspension outperforms the passive suspension in performance.
Key words:piezoelectric actuators;vehicle-road coupling;genetic algorithms;ride comfort

To address the issues of weakened target information and loss of fine details in fused images,this paper proposes a progressive multi-scale feature extraction and fusion method for infrared and visible image fusion.A structurally symmetric,parameter-independent dual-branch generation network is constructed.The original images and their enhanced versions are first fed into dilated convolution modules to extract contextual features at multiple scales,effectively capturing multi-scale information.Then,a multi-attention complementary residual aggregation module is introduced to enhance feature selectivity by emphasizing salient features and suppressing redundant ones,enabling progressive interaction and complementary fusion across scales.In the discriminator design,a dual-discriminator architecture is adopted to separately model the distributions of infrared and visible images,mitigating contrast shift and detail attenuation problems commonly encountered in multi-modal adversarial learning.Experimental results demonstrate that the proposed method outperforms most state-of-the-art algorithms in both objective metrics and subjective visual quality,retaining more texture details and achieving superior fusion performance.

Fire monitoring is crucial for reducing the loss of life and property.However,traditional methods suffer from insufficient real-time performance and accuracy in complex environments.This paper proposes a lightweight fire image detection algorithm based on improved YOLOv5s,which optimizes the monitoring system by integrating edge computing technology.By introducing the Convolutional Block Attention Module (CBAM) to enhance feature learning capabilities,employing Atrous Spatial Pyramid Pooling (ASPP) to expand the model’s receptive field,and utilizing the EIoU Loss function to accelerate convergence and improve regression accuracy,the improved model’s fire recognition rate is increased to 94%,with precision and recall rates reaching 94.2% and 92.4% respectively.By deploying the system on a modular AI module to directly process video data,cloud transmission latency is avoided,and detection real-time performance is significantly enhanced.This method provides an efficient solution for fire monitoring in complex scenarios and is of great significance for improving emergency response capabilities.

In this research,a strain was screened from the rotten kelp and was identified and named as Vibrio furnissii C1.Two alginate lyase genes alg792 and alg796 from V.furnissii C1 were heterologously expressed in Escherichia coli.The recombinant alginate lyases Alg792 and Alg796 were purified,and their properties were analyzed.The specific activities of Alg792 and Alg796 were 950 and 116 U/mg,respectively.The optimal temperature for Alg792 and Alg796 were 40 ℃ and 25 ℃,respectively.When incubated at 30 ℃,Alg792 possessed better enzyme viability stability.When the temperature was above 40 ℃,more than 80% of both enzymes activities were lost in 30 min.Both Alg792 and Alg796 possessed an optimum pH of 7.0.Alg796 possessed a wider pH applicable range,while Alg792 had better pH stability.Cu²⁺,Zn²⁺ and EDTA completely inhibited the activities of the two recombinases.Mg²⁺ and Na⁺ promoted the enzyme activities observably,while Ca²⁺,Mn²⁺ and K⁺ had opposite effects on the two recombinases.Both recombinant enzymes were bifunctional,Alg792 was endonuclease and Alg796 was exonuclease.These two alginate lyases have good application potential in the development of high value-added products of alginate.

Waste denim fabric,due to its resistance to degradation and low recycling efficiency,presents significant obstacles to effective resource reutilization,particularly as traditional recycling methods are often technologically complex and prone to causing deterioration in fiber performance,thereby limiting their reuse potential.In this study,waste denim was subjected to various pretreatment strategies and utilized as a reinforcing phase within a polylactic acid (PLA) matrix to prepare fabric-reinforced composites through a composite processing technique.The effects of different pretreatments,including alkali treatment and cyclic tensile conditioning,on the composites’ microstructure,mechanical behavior,and thermal properties were comprehensively investigated.The findings revealed that the combination of alkali treatment and cyclic tensile conditioning led to a marked enhancement in the overall mechanical performance of the composites.This research not only offers a novel route for the high-value reutilization of waste textiles but also contributes to the development of sustainable composite materials with promising application prospects in construction,automotive interiors,and aerospace engineering.