Multiple Experimental Sets Research Articles

RGISQL: Integrating Refined Grammatical Information into Relational Graph Neural Network for Text-to-SQL Task

The text-to-SQL task aims to convert natural language questions into corresponding SQL queries based on a given database schema. Previous models that rely on graph neural networks often struggle to accurately capture the complex grammatical relationships present in these questions, leading to poor performance when generating queries for longer requests. To address these challenges, we propose RGISQL, which integrates refined grammatical information extracted from the question and employs segmentation processing to effectively manage long queries. Additionally, RGISQL minimizes the complexity of edge embeddings by reducing the coupling within graph neural networks. By utilizing grammatical dependency trees, RGISQL is better equipped to capture the inherent structure and grammatical rules of questions. This refined grammatical information offers additional contextual and semantic cues for the model, thereby enhancing both its generalizability and interpretability. Furthermore, we dynamically assess the importance of different edges based on the graph structure, which helps reduce the coupling of edge embeddings and further improves the model’s performance. Multiple sets of experiments conducted on the Spider and Spider-Syn datasets demonstrate that RGISQL outperforms other baselines, achieving the best results in both datasets.

Electromagnetic signal analysis and recognition of coal and rock properties during the borehole pressure relief process: Simulation and experiment

Borehole pressure relief is an important measure to prevent and control rock bursts for high stress mine. However, the complex and changeable coal and rock properties cannot be effectively identified during the borehole pressure relief process, resulting in instability phenomena such as drill pipe jumping, drill sticking, and drill holding in the borehole system, which seriously affects the borehole pressure relief efficiency and operation safety. To perceive the drilling status of coal and rock in real time, this paper provides an in-depth analysis of the electromagnetic signals of coal and rock during drilling through simulations and experiments, so as to accurately identify the coal and rock properties. First, a small-size coaxially fed microstrip antenna that can be used on a specific drill pipe is designed, and some simulations are performed to analyze the propagation characteristics of electromagnetic signals under different drilling conditions. The electromagnetic signals of different coal and rock properties are significantly different in the range of 4-8 GHz, and the frequencies corresponding to the S11 extremes are 5.432 GHz, 5.436 GHz, 5.532 GHz, 5.552 GHz, 5.516 GHz, and 5.468 GHz, respectively. Then, the electromagnetic signals are mapped into time-frequency diagrams through continuous wavelet transform to provide training samples for coal and rock properties recognition model. Subsequently, the MobileNetV3-EACS model is proposed by embedding the efficient multi-channel attention mechanism and modifying the classification network structure to improve the detection accuracy and classification adaptability, achieving the effective recognition of coal and rock properties. Finally, an experimental platform of coal and rock borehole system is built and multiple sets of comparative experiments are conducted. The experimental results show that the MobileNetV3-EACS has a prediction accuracy of 99.11%, which is increased by 1.78% compared to the classical MobileNetV3, verifying the effectiveness and superiority of the proposed coal and rock properties recognition method.

A new framework for ultra-short-term electricity load forecasting model using IVMD–SGMD two–layer decomposition and INGO–BiLSTM–TPA–TCN

Accurate forecasting of electricity loads is crucial for the development of electricity scheduling and supply services. With the increase in distributed electricity energy sources and the complexity of electricity systems, the strong volatility of load changes brings more challenges to the reliability of load forecasting. Therefore, we propose a novel ultra-short-term electricity load forecasting model using improved two-layer decomposition and an improved deep learning model with temporal pattern attention based on improved northern goshawk optimization (INGO). First, the load data were decomposed thoroughly using the two-layer decomposition model of improved variational mode decomposition (IVMD) with parameter optimization by INGO and symplectic geometry mode decomposition (SGMD) to improve the interpretability of the subsequences. Subsequently, INGO is used to optimize the parameters in bidirectional long short-term memory (BiLSTM). Temporal pattern attention (TPA) is added to BiLSTM, which extracts complex relationships from the hidden neurons of BiLSTM and selects relevant information from different time scales. After predicting and reconstructing the subsequences, the temporal convolutional network (TCN) prediction model is used to perform error correction to improve the final prediction accuracy. Because many government reports and policy information are summarized and published quarterly, to provide information support, we divide the electricity load datasets of two countries by quarters before forecasting. By performing multiple sets of experiments on the two datasets, it is demonstrated that the proposed model has high precision and robustness, and the obtained ultra-short-term electricity load forecasting results can accurately fit the load fluctuation trend.

Multimodal Fusion Anomaly Detection Model for Agricultural Wireless Sensors

ABSTRACTAgricultural Internet of Things has become one of the most important data sources of agricultural big data. However, the wireless sensors equipped with agricultural Internet of Things is affected by many factors, and anomalous data inevitably exists during the data collection process. The existence of anomalous data leads to the deterioration of the agricultural data quality, which hinders the efficient development of agricultural data analysis. In order to detect anomalous data accurately for agricultural wireless sensors, in this article, we propose an anomaly detection model that combines multimodal fusion and error reconstruction. The model first adds Gaussian noise to the input sequence and converts it into standard time series and image data. Subsequently, it is processed by different modal encoders and fuses the image and sequence modal data using Temporal Cross‐modal Attention module to enhance the perception of anomalous modal information. Finally, the fused data is reconstructed, and anomalous data are identified by the image and sequence decoders. Comparison experiments with several baselines prove the validity of the model proposed in this article, ablation experiments prove the necessity of the key modules in the model, and multiple sets of experiments are also designed to discuss the effects of hyperparameters.

Intelligent aeration amount prediction control for wastewater treatment process based on recurrent neural network

The use of machine learning in artificial intelligence to solve industrial problems has been a current trend. Predictions can be made by machine learning algorithms. This greatly improves the efficiency and also the accuracy. With the development of industrialization, the pollution of water resources is becoming more and more serious. How to treat wastewater more cost-effectively to meet the discharge standards has become one of the most urgent problems in the world. In the Anaerobic-Anoxic-Aerobic (AAO) process, the key to reach the standard of sewage treatment with low cost and high efficiency lies in the aeration link. In this study, multiple machine learning methods are used for modeling, and proposed a method employing the long short-term memory (LSTM) neural network model for regression prediction addresses the need for a more accurate prediction approach. By utilizing data from the preceding 100 days, this model replaces manual, experience-based prediction methods, thereby mitigating energy consumption. This model is optimized by training and testing with real wastewater treatment plant data and continuously adjusting the parameters through multiple sets of experiments. This can make the real recorded aeration amount and the actual aeration amount infinitely close. Through the experimental comparison, it is found that the performance of the LSTM neural network model is better, with about 14% higher accuracy than the benchmark model.

Workflow scheduling based on asynchronous advantage actor–critic algorithm in multi-cloud environment

Recently, the multi-cloud environment (MCE) has increasingly become the preferred choice of users. As with the cloud environment, efficient workflow scheduling in a MCE remains crucial for identifying the cost efficiency and overall performance of the MCE. In MCE, the resources exhibit heterogeneity, complexity, and dynamism. Simultaneously, the intricate inter-task dependencies among workflow tasks, diverse Quality of Service (QoS) metrics for users, and multiple cloud service providers’ (CSPs) billing mechanisms significantly amplify the workflow scheduling challenge. Motivated by the application of reinforcement learning (RL) in workflow scheduling in a cloud environment, this paper proposes a scheduling algorithm that takes advantage of the asynchronous advantage actor–critic algorithm (A3C) to balance cost, makespan and resource utilization in workflow scheduling in a MCE. By analyzing the elements in the MCE, we design and define multiple agents in the MCE, and each cloud service provider will have an agent to record the state and update the local parameters. For the workflow task submitted by the user, the action is selected according to the initialization policy and submitted to the scheduling action to allocate the task to a designated virtual machine in the MCE so that each agent can more clearly perceive the environment change and adapt to the MCE. In contrast to the traditional A3C algorithm, we design a new critic network according to the data characteristics of real-world scientific workflows so that each agent is more suitable for real-world scientific workflow data. Through multiple sets of simulation experiments, the workflow scheduling algorithm based on the A3C algorithm in the MCE (MCWS-A3C) was compared with three benchmark methods. The experimental results show that the proposed method has better advantages than other methods in terms of cost, makespan, and resource utilization. Specifically, on the Montage_100 dataset, the average cost was reduced by 55.12% compared to other methods. The pioneering introduction of the A3C algorithm that adapts to the dynamic environment into the MCE brings more possibilities to address the issue of workflow scheduling in the MCE.

Path planning for Multi-USV target coverage in complex environments

This research introduces a path planning methodology aimed at coordinating multiple unmanned surface vehicles (USVs) to accomplish target coverage task in obstacle-rich environments. Leveraging the locking sweeping method (LSM), our proposed method constructs task target point distance fields that integrates environmental constraints, alongside a task target point distance matrix. During the task allocation phase, we design a cost assessment function to assess the generated allocation solutions. And a greedy allocation strategy is employed to determine the optimal number of USVs required for mission completion, ensuring evenly distribution of target task points among them. Subsequently, we improve the ant colony optimization (ACO) method by redesigning the heuristic function and pheromone update rules, considering environmental constraints and task execution sequence constraints. This refinement facilitates the generation of optimized task execution sequences and safe navigation paths in obstacle environments. The effectiveness of the proposed method is validated through multiple sets of simulation experiments and compared with existing methods. The results demonstrate the practicality and efficacy of the method in addressing the challenges of USVs target coverage task in obstacle environments.

Closed-loop transfer enables artificial intelligence to yield chemical knowledge.

Artificial intelligence-guided closed-loop experimentation has emerged as a promising method for optimization of objective functions1,2, but the substantial potential of this traditionally black-box approach to uncovering new chemical knowledge has remained largely untapped. Here we report the integration of closed-loop experiments with physics-based feature selection and supervised learning, denoted as closed-loop transfer (CLT), to yield chemical insights in parallel with optimization of objective functions. CLT was used to examine the factors dictating the photostability in solution of light-harvesting donor-acceptor molecules used in a variety of organic electronics applications, and showed fundamental insights including the importance of high-energy regions of the triplet state manifold. This was possible following automated modular synthesis and experimental characterization of only around 1.5% of the theoretical chemical space. This physics-informed model for photostability was strengthened using multiple experimental test sets and validated by tuning the triplet excited-state energy of the solvent to break out of the observed plateau in the closed-loop photostability optimization process. Further applications of CLT to additional materials systems support the generalizability of this strategy for augmenting closed-loop strategies. Broadly, these findings show that combining interpretable supervised learning models and physics-based features with closed-loop discovery processes can rapidly provide fundamental chemical insights.

A Self-Learning Hyper-Heuristic Algorithm Based on a Genetic Algorithm: A Case Study on Prefabricated Modular Cabin Unit Logistics Scheduling in a Cruise Ship Manufacturer.

Hyper-heuristic algorithms are known for their flexibility and efficiency, making them suitable for solving engineering optimization problems with complex constraints. This paper introduces a self-learning hyper-heuristic algorithm based on a genetic algorithm (GA-SLHH) designed to tackle the logistics scheduling problem of prefabricated modular cabin units (PMCUs) in cruise ships. This problem can be regarded as a multi-objective fuzzy logistics collaborative scheduling problem. Hyper-heuristic algorithms effectively avoid the extensive evaluation and repair of infeasible solutions during the iterative process, which is a common issue in meta-heuristic algorithms. The GA-SLHH employs a genetic algorithm combined with a self-learning strategy as its high-level strategy (HLS), optimizing low-level heuristics (LLHs) while uncovering potential relationships between adjacent decision-making stages. LLHs utilize classic scheduling rules as solution support. Multiple sets of numerical experiments demonstrate that the GA-SLHH exhibits a stronger comprehensive optimization ability and stability when solving this problem. Finally, the validity of the GA-SLHH in addressing real-world decision-making issues in cruise ship manufacturing companies is validated through practical enterprise cases. The results of a practical enterprise case show that the scheme solved using the proposed GA-SLHH can reduce the transportation time by up to 37%.

Design of nonlinear locking mechanism for shape memory alloy archwire of miniature orthodontic device

The locking mechanism between bracket and shape memory alloy (SMA) archwire in the newly developed domestic orthodontic device is the key to controlling the precise alignment of the teeth. To meet the demand of locking force in clinical treatment, the tightening torque angle of the locking bolt and the required torque magnitude need to be precisely designed. For this purpose, a design study of the locking mechanism is carried out to analyze the correspondence between the tightening torque angle and the locking force and to determine the effective torque value, which involves complex coupling of contact, material and geometric nonlinear characteristics. Firstly, a simulation analysis based on parametric orthogonal experimental design is carried out to determine the SMA hyperelastic material parameters for the experimental data of SMA archwire with three-point bending. Secondly, a two-stage fine finite-element simulation model for bolt tightening and archwire pulling is established, and the nonlinear analysis is converged through the optimization of key contact parameters. Finally, multiple sets of calibration experiments are carried out for three tightening torsion angles. The comparison results between the design analysis and the calibration experiments show that the deviation between the design analysis and the calibration mean value of the locking force in each case is within 10%, and the design analysis method is valid and reliable. The final tightening torque angle for clinical application is determined to be 10° and the rated torque is 2.8 N∙mm. The key data obtained can be used in the design of clinical protocols and subsequent mechanical optimization of novel orthodontic devices, and the research methodology can provide a valuable reference for force analysis of medical devices containing SMA materials.

A transfer learning model for cognitive electronic reconnaissance of unmanned aerial vehicle: Experiments

Applying Deep Reinforcement Learning (DRL) technologies to Unmanned Aerial Vehicle (UAV) electronic reconnaissance is one of the current research hotspots. However, simulation and engineering practice show that due to poor generalization of DRL models, the performance of Cognitive Electronic Reconnaissance (CER) policies based on training will significantly decrease when the mission scene undergoes slight changes. To address this issue, we thoughtfully combine the mission area segmentation technique with transfer DRL and propose a difference-adaptive transfer DRL algorithm. This algorithm involves mission subarea segmentation, subarea pre-training, multi-subarea policy transfer, and multi-subarea splicing, providing an efficient solution to the convergence problem of the DRL algorithm caused by mission space expansion and reward sparsity. Additionally, a general CER transfer learning simulator is developed based on the analysis of the capabilities of the maneuvering platform and electronic reconnaissance payload. Multiple sets of CER policy transfer learning experiments are designed for different mission spaces, mission difficulties, and UAV characteristics. Compared with the algorithm baseline, our designed policy model significantly outperforms: the mission completion rate of UAVs in multi-scale mission spaces improves by up to 37.4%, reaching 97.5%, while the training time is reduced by 2.46 h. Further behavior analysis shows that this policy model enables UAVs to exhibit target tracking behaviors such as hovering and sustained approach.

Enhancing the performance of PVT-TEG power generation systems by heat pipes and Fe3O4 nanofluids

In recent years, enhancing the cooling of photovoltaic (PV) panels to improve overall energy efficiency has become a significant focus. In this study, we propose an improved power generation system integrating semiconductor thermoelectric generators (TEGs). The integration involves the installation of several heat pipes and serpentine copper tubes on the back of PV panels, with nanofluids flowing through the copper tubes as the working fluid, effectively cooling the PV panels. The copper tube is connected to the heat exchanger on the high-temperature side of the TEGs, forming a closed-loop circulation system. Consequently, the cooling fluid transfers heat to the hot surface of the TEGs, utilizing waste heat generated by the PV panels for electricity generation. The TEGs consist of 10 thermoelectric modules, with the low-temperature end in contact with the finned aluminum plate (TEGs radiator) and actively cooled by natural air cooling. Additionally, to validate the design rationale, multiple sets of comparative experiments were arranged. The experiments utilized nanofluids with volume fractions of 0.5 %, 1 %, 2 %, and 3 %, and water, as the cooling fluid. The experimental results demonstrated that the maximum operating temperature of the reference PV panel can reach 81.9 °C, and the average and maximum exergy efficiencies throughout a day are recorded at 16.91 % and 31.84 %, respectively. It’s noteworthy that the most significant cooling effect occurs when the nanofluid volume fraction is 2 %, reducing the maximum temperature of the PVT panel to 61.5 °C. Simultaneously, the average daily exergy efficiency and exergy efficiency of the PVT-TEG power generation system reached peak values of 21.06 % and 39.87 %, respectively. Compared to the reference PV panels, the output power of the PVT-TEG system increases by up to 30.51 %.

Plug-and-play learnable EMD module for time-domain speech separation

This paper presents a learnable empirical mode decomposition (EMD) module tailored for single channel speech separation. The proposed EMD module exhibits a flexible plug-and-play characteristic, enabling its integration into various speech separation networks. Conventionally, non-learnable EMD algorithms, as a nonlinear time-domain adaptive techniques, aimed at decomposing speech signals into a set of oscillatory intrinsic mode functions (IMFs) and a residual component. However, the inherent non-differentiability of the interpolation operation in these algorithms restricts their integration into neural networks, thus the EMD algorithm only exists in the preprocessing step and cannot be integrated into the neural network to form an end-to-end trainable speech separation network. In order to mitigate this issue, we use the max pooling layer and the deconvolution network to replace the envelope solving algorithm in the EMD method and build a trainable EMD neural network module. Multiple sets of experiments have shown that the integrating of EMD module brings the improvement to different separation systems.

Exploring the predictive ability of the CA–Markov model for urban functional area in Nanjing old city

With advancements in sustainable urban development, research on urban functional areas has garnered significant attention. In recent years, Point-of-Interest, with their large volume of information and ease of acquisition, have been widely applied in research on urban functional domains. However, scholars currently focus on the identification of urban functional areas, usually relying on data from a single period, whereas research on the prediction of functional areas has not yet been well validated. Therefore, in this study, we propose a new method based on several years of POI data to predict urban functional areas. Taking Nanjing City, Jiangsu Province, as an example, we first identified the functional area distribution of the old city of Nanjing over several years using POI data and then designed multiple sets of experiments to explore the CA–Markov model’s ability to predict functional areas from various aspects, including model overall accuracy, robustness, and comparison analysis between predictions and actual situations. The results show that (1) for mixed or single functional areas, the model’s predictions over several years tend to be stable, and the accuracy of the predictions over many years indicates the robustness of the model in predicting urban functional areas. (2) For mixed functional areas in cities, model predictions largely rely on the distribution of the base years used for prediction, leading to inaccurate results; thus, it is still not applicable for simulating and predicting mixed functional areas. (3) For single functional areas in cities or primary functions within an area, the model’s predicted degree of change was close to the actual degree of change, making the results referable.

A Multi-Objective Non-Dominated Sorting Gravitational Search Algorithm for Assembly Flow-Shop Scheduling of Marine Prefabricated Cabins

Prefabricated cabin modular units (PMCUs) are a widespread type of intermediate products used during ship or offshore platform construction. This paper focuses on the scheduling problem of PMCU assembly flow shops, which is summarized as a multi-objective, fuzzy-blocking hybrid flow-shop-scheduling problem based on learning and fatigue effects (FB-HFSP-LF) to minimize the maximum fuzzy makespan and maximize the average fuzzy due-date agreement index. This paper proposes a multi-objective non-dominated sorting gravitational search algorithm (MONSGSA) to solve it. In the proposed MONSGSA, the ranked-order value is used to convert continuous solutions to discrete solutions. Multi-dimensional Latin hypercube sampling is used to enhance initial population diversity. Setting up an external archive to maintain non-dominated solutions while introducing an adaptive inertia factor and a trap avoidance operator to guide individual positional updates. The results of multiple sets of experiments show that Pareto solutions of MONSGSA have better distribution and convergence compared to other competitors. Finally, the instance of PMCU manufacturer is used for validation, and the results show that MONSGSA has better applicability to practical problems.

Mapping model of ribbon contour and tool influence function based on distributed parallel neural networks in magneto-rheological finishing.

Magnetorheological finishing (MRF) stands out as a notable polishing technology, characterized by high precision and minimal damage. However, establishing an accurate and practical model for the tool influence function (TIF) of MRF poses a significant challenge. In this paper, a TIF modeling method of MRF based on distributed parallel neural networks is proposed for the first time. Assessment of the viability of this approach through multiple sets of robot-assisted MRF experiments is detailed. The experimental results conclusively demonstrate the successful intelligent prediction of TIF, with key indicators such as volume removal rate and peak removal rate achieving an average prediction accuracy exceeding 95%. This method can remarkably advance the intelligence of the TIF model in MRF and serve as a valuable reference for other optical fabrication methods.

Filter bank temporally local multivariate synchronization index for SSVEP-based BCI

BackgroundMultivariate synchronization index (MSI) has been successfully applied for frequency detection in steady state visual evoked potential (SSVEP) based brain–computer interface (BCI) systems. However, the standard MSI algorithm and its variants cannot simultaneously take full advantage of the time-local structure and the harmonic components in SSVEP signals, which are both crucial for frequency detection performance. To overcome the limitation, we propose a novel filter bank temporally local MSI (FBTMSI) algorithm to further improve SSVEP frequency detection accuracy. The method explicitly utilizes the temporal information of signal for covariance matrix estimation and employs filter bank decomposition to exploits SSVEP-related harmonic components.ResultsWe employed the cross-validation strategy on the public Benchmark dataset to optimize the parameters and evaluate the performance of the FBTMSI algorithm. Experimental results show that FBTMSI outperforms the standard MSI, temporally local MSI (TMSI) and filter bank driven MSI (FBMSI) algorithms across multiple experimental settings. In the case of data length of one second, the average accuracy of FBTMSI is 9.85% and 3.15% higher than that of the FBMSI and the TMSI, respectively.ConclusionsThe promising results demonstrate the effectiveness of the FBTMSI algorithm for frequency recognition and show its potential in SSVEP-based BCI applications.

BearingFM: Towards a foundation model for bearing fault diagnosis by domain knowledge and contrastive learning

Monitoring bearing failures in production equipment can effectively prevent finished product quality issues and unplanned factory downtime, thereby reducing supply chain uncertainties and risk. Therefore, monitoring bearing failures in production equipment is important for improving supply chain sustainability. Due to the generalization limitations of neural network models, specific models must be trained for specific tasks. However, in real industrial scenarios, there is a severe lack of labeled samples, making it difficult to deploy fault diagnosis models across massive amounts of equipment in workshops. In order to solve the above issue, this paper proposes a cloud-edge-end collaborative semi-supervised learning framework, which provides multi-level computing power and data support for building a foundation model. A data augmentation method based on the bearing fault mechanism is proposed, which effectively preserves the inherent essential characteristics in vibration signals by normalizing frequency and adding noise in specific frequency bands. A novel contrastive learning model is designed, which narrows the distances between positive samples and widens the distances between negative samples in the high-dimensional space through cross comparisons in the time dimension and knowledge dimension, thereby extracting the most essential characteristics from the unlabeled signals. Multiple sets of experiments conducted on four datasets demonstrate that the proposed approach achieves an approximately 98% fault classification accuracy with only 1.2% labeled samples.

Machining quality prediction of multi-feature parts using integrated multi-source domain dynamic adaptive transfer learning

Machining quality prediction of multi-feature parts has been a challenging problem because of small dataset and inconsistent quality data distribution with respect to each machining feature. Transfer learning that leverages knowledge of one task and can be repurposed on another task seems a good solution for this purpose. However, traditional transfer learning typically has a single source domain and a target domain, which limits its applications in multi-source scenarios (e.g., multi-feature). To solve this issue, this paper proposes a novel integrated multi-source domain dynamic adaptive transfer learning (IMD-DATL) framework for machining quality prediction of multi-feature part machining systems. Specifically, a domain-sample similarity double matching multi-source domain integration method is designed to construct the integration knowledge transfer from multiple source domains to the target domain. A residual feature extraction network based on sample entropy-dynamic channel double-layer attention structure and a fine-grained transferable feature attention module are designed. These three attentions are used to improve the feature learning ability and the adaptation level to the predicted object in the three dimensions of sample, channel and data feature. Finally, multiple sets of comparative experiments in thin-walled part machining systems confirm the effectiveness and superiority of the proposed method for cross-domain quality prediction. Compared with other traditional transfer learning methods, the MAE, RMSE and Score on average of this method are increased by 5.47 %, 4.59 % and 4.84 %, respectively, compared with other multi-source domain adaptation methods, the MAE, RMSE and Score on average of this method are increased by 7.13 %, 7.37 % and 6.52 %, respectively.

Residual Temporal Convolutional Network With Dual Attention Mechanism for Multilead-Time Interpretable Runoff Forecasting.

As a pivotal subfield within the domain of time series forecasting, runoff forecasting plays a crucial role in water resource management and scheduling. Recent advancements in the application of artificial neural networks (ANNs) and attention mechanisms have markedly enhanced the accuracy of runoff forecasting models. This article introduces an innovative hybrid model, ResTCN-DAM, which synergizes the strengths of deep residual network (ResNet), temporal convolutional networks (TCNs), and dual attention mechanisms (DAMs). The proposed ResTCN-DAM is designed to leverage the unique attributes of these three modules: TCN has outstanding capability to process time series data in parallel. By combining with modified ResNet, multiple TCN layers can be densely stacked to capture more hidden information in the temporal dimension. DAM module adeptly captures the interdependencies within both temporal and feature dimensions, adeptly accentuating relevant time steps/features while diminishing less significant ones with minimal computational cost. Furthermore, the snapshot ensemble method is able to obtain the effect of training multiple models through one single training process, which ensures the accuracy and robustness of the forecasts. The deep integration and collaborative cooperation of these modules comprehensively enhance the model's forecasting capability from various perspectives. Ablation studies conducted validate the efficacy of each module, and through multiple sets of comparative experiments, it is shown that the proposed ResTCN-DAM has exceptional and consistent performance across varying lead times. We also employ visualization techniques to display heatmaps of the model's weights, thereby enhancing the interpretability of the model. When compared with the prevailing neural network-based runoff forecasting models, ResTCN-DAM exhibits state-of-the-art accuracy, temporal robustness, and interpretability, positioning it at the forefront of contemporary research.