Single-step Prediction Research Articles

Relevance-Based Reconstruction Using an Empirical Mode Decomposition Informer for Lithium-Ion Battery Surface-Temperature Prediction

Accurate monitoring of lithium-ion battery temperature is essential to ensure these batteries’ efficient and safe operation. This paper proposes a relevance-based reconstruction-oriented EMD-Informer machine learning model, which combines empirical mode decomposition (EMD) and the Informer framework to estimate the surface temperature of 18,650 lithium-ion batteries during charging and discharging processes under complex operating conditions. Initially, based on 9000 data points from the U.S. NASA Prognostics Center of Excellence’s random battery-usage dataset, where each data point includes three features: temperature, voltage, and current, EMD is used to decompose the temperature data into intrinsic mode functions (IMFs). Subsequently, the IMFs are reconstructed into low-, medium-, and high-correlation components based on their correlation with the original data. These components, along with voltage and current data, are fed into sub-models. Finally, the model captures the long-term dependencies among temperature, voltage, and current. The experimental results show that, in single-step prediction, the mean squared error, mean absolute error, and maximum absolute error of the model’s predictions are 0.00095, 0.02114, and 0.32164 °C; these metrics indicate the accurate prediction of the surface temperature of lithium-ion batteries. In multi-step predictions, when the prediction horizon is set to 12 steps, the model achieves a hit rate of 93.57% where the maximum absolute error is within 0.5 °C; under these conditions, the model combines high predictive accuracy with a broad predictive range, which is conducive to the effective prevention of thermal runaway in lithium-ion batteries.

Deep learning in template-free de novo biosynthetic pathway design of natural products.

Natural products (NPs) are indispensable in drug development, particularly in combating infections, cancer, and neurodegenerative diseases. However, their limited availability poses significant challenges. Template-free de novo biosynthetic pathway design provides a strategic solution for NP production, with deep learning standing out as a powerful tool in this domain. This review delves into state-of-the-art deep learning algorithms in NP biosynthesis pathway design. It provides an in-depth discussion of databases like Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, and UniProt, which are essential for model training, along with chemical databases such as Reaxys, SciFinder, and PubChem for transfer learning to expand models' understanding of the broader chemical space. It evaluates the potential and challenges of sequence-to-sequence and graph-to-graph translation models for accurate single-step prediction. Additionally, it discusses search algorithms for multistep prediction and deep learning algorithms for predicting enzyme function. The review also highlights the pivotal role of deep learning in improving catalytic efficiency through enzyme engineering, which is essential for enhancing NP production. Moreover, it examines the application of large language models in pathway design, enzyme discovery, and enzyme engineering. Finally, it addresses the challenges and prospects associated with template-free approaches, offering insights into potential advancements in NP biosynthesis pathway design.

Multi-step prediction of offshore wind power based on Transformer network and Huber loss

In the context of the burgeoning expansion of renewable energy sources, the precise prediction of offshore wind power assumes a pivotal role in safeguarding the reliability, economic viability, and sustainable progression of offshore wind farms. The present study introduces a novel methodology for offshore wind power prediction, predicated upon the synergy of the Transformer network and Huber loss function. Empirical validation is conducted utilizing authentic data from a European offshore wind farm. The resulting analyses delineate a discernible superiority of the Transformer network over classical LSTM and GRU models in capturing the intricate long-term dependencies intrinsic to the time series. Furthermore, the inclusion of the Huber loss function effectively mitigates the challenges posed by the high volatility often characteristic of offshore wind power data. The study also demonstrates the beneficial integration of autoencoder reconstruction for denoising and slime mould optimization algorithm to augment prediction performance. Distinctively diverging from traditional single-step prediction paradigms, the multi-step prediction model constructed within this research offers a more comprehensive and precise prediction of wind power. Such an innovative approach represents a valuable contribution to the field, with tangible implications for the dependable operation and future advancement of wind power.

Multi-step electric vehicles charging loads forecasting: An autoformer variant with feature extraction, frequency enhancement, and error correction blocks

The extensive integration of electric vehicles (EVs) into power grids has raised safety concerns. Meanwhile current EVs charging loads forecasting models mainly focus on single-step predictions, neglect load sparsity and underutilize residuals, limiting their accuracy. To tackle these issues, a multi-step short-term forecasting model, specifically designed to predict charging loads at public EVs charging stations using historical data, is proposed. The model begins with a preprocessing phase to handle missing data and outliers. Subsequently, an Autoformer variant is intruduced, which initially integrates feature extraction block leveraging the Temporal Convolutional Network (TCN) to adeptly capture temporal dynamics. This is further enhanced with a Discrete Cosine Transform (DCT)-based frequency enhancement block to learn sparsity and periodicity features of EVs charging loads by frequency mapping. In addition, an advanced error correction block, comprising Bayesian optimization algorithm, Convolutional Neural Networks, Multi-Head Self-Attention, and Gated Recurrent Units (BCMG), is incorporated to refine initial prediction results. Above all, the TCN-DCT enhanced Autoformer-BCMG (TDformerBCMG) model is proposed. Finally, a comprehensive experimental analysis is conducted on ACN-data for EVs charging load forecasting in 1–3 h steps. It is demonstrated that TDformerBCMG not only is effective and superior in each component, but also achieves a reduction in Mean Absolute Error by 20.94 % to 74.05 % compared to existing models. Moreover, stability is evidenced across 50 repeated trials, with the standard deviation for the Coefficient of Determination varying between 8.5844e-3 and 3.1580e-2. Further discussion on application analysis is provided. Consequently, TDformerBCMG offers a viable methodology and establishes a new benchmark for power system implementation.

3DTCN-CBAM-LSTM short-term power multi-step prediction model for offshore wind power based on data space and multi-field cluster spatio-temporal correlation

The accuracy of offshore wind power forecasting (OWPF) is the basis for guaranteeing the safe dispatch and economic operation of power systems, which can reduce the technical and economic risks faced by power market participants. Combining the theoretical approaches of data space and deep learning, this study proposes a hybrid short-term OWPF framework based on the integrated consideration of spatial and temporal correlations of regional field clusters in the data space and the 3DTCN-CBAM-LSTM model. Firstly, other offshore wind farms with spatio-temporal correlation with the target offshore wind farm are screened out using the K-means spatial clustering algorithm. The historical power sequences determined to be similar to the target offshore wind farm and feature data of target offshore wind farms are used as inputs to the prediction model to improve the efficiency of the framework's power prediction. Secondly, the short-term power prediction framework of offshore wind power based on the 3DTCN-CBAM-LSTM model is constructed, using the 3DTCN model's ability to capture the spatio-temporal dynamic characteristics of offshore wind farm data to improve the accuracy of the prediction model. The temporal characteristics of the data are strengthened using the LSTM model to improve the stability of the prediction model. Finally, the CBAM mechanism is introduced to combine channel attention and spatial attention to eliminate irrelevant information and capture the spatio-temporal characteristics of the data in a more targeted way, which further improves the accuracy of OWPF. To verify the practicality and reliability of the wind power prediction framework described in this study, the differences between the prediction results of the framework and those of the benchmark model are compared using single-step prediction and multi-step prediction (4, 6, 8, and 12 steps). Compared with the benchmark model (BPNN), the single-step prediction results show that the MAE, MAPE, and MSE metrics of the 3DTCN-CBAM-LSTM model are reduced by 20.004 MW, 50.13%, 55.61%, and 80.29%, respectively, and the R2 is improved by 3.39%, which is a substantial improvement in prediction performance. The multi-step prediction results show that, with the progression of the prediction steps, the 3DTCN-CBAM-LSTM model has the highest prediction accuracy and model fit at each step. Therefore, the proposed 3DTCN-CBAM-LSTM model presents obvious advantages and reliability in the prediction of offshore wind power multi-step power.

1D -FNO - Residual convolutional neural network autoregressive prediction model for deep water wave train evolution based on high-order spectral (HOS) data

Accurate prediction of deep-water wave dynamics is essential for ocean engineering, meteorology, and maritime safety. Traditional physical models, though effective, face significant computational and time challenges with complex nonlinear dynamics. This study introduces an autoregressive prediction model using a one-dimensional adaptive Fourier Neural Operator-Residual Convolutional Neural Network (1D-FNO-ResNet) to enhance prediction accuracy. We developed a high-precision single-step prediction model capable of capturing wave characteristics such as height and fluid particle velocity. A fine-tuning training strategy based on multi-step prediction errors was proposed for autoregressive evolution. Performance comparisons between the ResNet and FNO-ResNet models were made under various fine-tuning steps. A correction model was also introduced to control error accumulation, improving autoregressive prediction steps. The model's generalization performance was validated with modulated and random wave trains under different conditions, optimizing predictive stability and accuracy. This approach demonstrates efficient 1D deep-water wave evolution prediction, offering a novel deep learning solution for wave dynamics prediction in ocean engineering, supporting wave energy conversion device design and optimization.

AnimalEnvNet: A Deep Reinforcement Learning Method for Constructing Animal Agents Using Multimodal Data Fusion

Simulating animal movement has long been a central focus of study in the area of wildlife behaviour studies. Conventional modelling methods have difficulties in accurately representing changes over time and space in the data, and they generally do not effectively use telemetry data. Thus, this paper introduces a new and innovative deep reinforcement learning technique known as AnimalEnvNet. This approach combines historical trajectory data and remote sensing images to create an animal agent using deep reinforcement learning techniques. It overcomes the constraints of conventional modelling approaches. We selected pandas as the subject of our research and carried out research using GPS trajectory data, Google Earth images, and Sentinel-2A remote sensing images. The experimental findings indicate that AnimalEnvNet reaches convergence during supervised learning training, attaining a minimal mean absolute error (MAE) of 28.4 m in single-step prediction when compared to actual trajectories. During reinforcement learning training, the agent has the capability to replicate animal locomotion for a maximum of 12 iterations, while maintaining an error margin of 1000 m. This offers a novel approach and viewpoint for mimicking animal behaviour.

Multiscale graph based spatio-temporal graph convolutional network for energy consumption prediction of natural gas transmission process

Energy consumption prediction is crucial for planning and managing natural gas networks, enhancing transmission efficiency. However, most of relevant studies fail to consider the topology of natural gas network. Therefore, a multiscale graph based spatio-temporal graph convolutional network model (MG-STGCN) is proposed, which can extract the multi-scale relationship of data. Based on graph convolutional and temporal gated convolution network, data correlation, spatial dependence and time dependence are considered to improve the prediction effect. In order to verify the effect of the model, the model is tested based on two natural gas networks in different region of China. The single step prediction results of MG-STGCN are significantly better than spatio-temporal graph convolutional network (STGCN), long short-term memory (LSTM), gate recurrent unit (GRU) and temporal convolutional network (TCN), which proves the effectiveness and generalization ability of MG-STGCN. The training time of MG-STGCN is reduced by 20.8 % compared with STGCN. In different prediction time step, the average prediction effect of MG-STGCN is better than that of other methods, and the performance improvement rate is not less than 8.6 %. MG-STGCN can well capture the parameter correlation, spatial and temporal dependence of data, and achieve effective prediction of pipeline energy consumption.

A spatial–temporal data-driven deep learning framework for enhancing ultra-short-term prediction of distributed photovoltaic power generation

Effective utilization of spatial–temporal information can improve the accuracy of ultra-short-term prediction of power generation from distributed photovoltaic (PV) stations in the region. This paper introduces an ultra-short-term spatial–temporal prediction model for distributed PV power generation, blending data-driven methodology with deep learning technique. The model integrates a self-attention mechanism (SA), a Bi-directional Convolutional Gated Recurrent Unit (BiConvGRU), and an encoder-decoder structure, called ABCGRU. The spatial–temporal attributes of PV power generation can be effectively utilized to accurately predict the output of PV power stations at different locations. Firstly, this paper proposes a 2D distributed PV measurement frame approach considering the spatial–temporal properties of PV power. The combination of Pearson correlation coefficient, the normalized Euclidean distance, the Shape-based distance (SBD) analysis based on cross-correlation and geographic distance reduces the input dimensionality. Secondly, to better capture the spatial–temporal patterns within the 2D distributed PV measurement frame, this paper proposes the ABCGRU model. Finally, the predictive performance of the model is verified through experiments. On the Birmingham dataset, the relative absolute error (RAE) for single-step (15 min) prediction is 0.13, and the average RAE for multi-step (30–60 min) prediction is about four times higher than ConvGRU. The single-step prediction RAE of Little Rock and New Orleans datasets is about 3–4 times higher than ConvGRU. In the comparison between the same series of models, the 4-layer ABCGRU has the highest accuracy. Moreover, the effectiveness of data dimensionality reduction was verified through experimental comparison. The RAE for single-step prediction on the Datong dataset is 0.0048.

VTion-PatchTST: Elevated PatchTST model for network security situation prediction

The rapid development of network attack techniques increases the risk of the internet being attacked, thus necessitating timely and effective network protection mechanisms. Network security situation prediction is a proactive defense method, the results of which can help formulate defense strategies in advance. The current network security situation prediction mainly focuses on single-step prediction, and the accuracy of multi-step prediction needs to be improved. By introducing three technologies: Variational Mode Decomposition (VMD), Temporal Convolutional Networks (TCN), Feature Fusion, an improved PatchTST model is proposed to solute the issue. VMD decomposes a complex sequence of historical situational values into multiple relatively stable subsequences with different features to make the model channel modeling specific. TCN enhances the ability of the model to extract temporal features. Feature fusion makes predictions more comprehensive through cross-channel linking. Both experiments and evaluations are conducted on the UNSW-NB15 and CIC-IDS2017 public datasets, and nine baseline models are used to compare with the proposed VTion-PatchTST. The experimental results show that the proposed model is more applicable to cybersecurity situation forecasting, with an relative average reduction of 40.3% and 29.0% in Mean Square Error (MSE) and Mean Absolute Error (MAE), respectively.

ST-D3DDARN: Urban traffic flow prediction based on spatio-temporal decoupled 3D DenseNet with attention ResNet.

Urban traffic flow prediction plays a crucial role in intelligent transportation systems (ITS), which can enhance traffic efficiency and ensure public safety. However, predicting urban traffic flow faces numerous challenges, such as intricate temporal dependencies, spatial correlations, and the influence of external factors. Existing research methods cannot fully capture the complex spatio-temporal dependence of traffic flow. Inspired by video analysis in computer vision, we represent traffic flow as traffic frames and propose an end-to-end urban traffic flow prediction model named Spatio-temporal Decoupled 3D DenseNet with Attention ResNet (ST-D3DDARN). Specifically, this model extracts multi-source traffic flow features through closeness, period, trend, and external factor branches. Subsequently, it dynamically establishes global spatio-temporal correlations by integrating spatial self-attention and coordinate attention in a residual network, accurately predicting the inflow and outflow of traffic throughout the city. In order to evaluate the effectiveness of the ST-D3DDARN model, experiments are carried out on two publicly available real-world datasets. The results indicate that ST-D3DDARN outperforms existing models in terms of single-step prediction, multi-step prediction, and efficiency.

Short-term water demand prediction based on decomposition technique optimization and a multihead attention mechanism

ABSTRACT Short-term water demand prediction is crucial for real-time optimal scheduling and leakage control in water distribution systems. This paper proposes a new deep learning-based method for short-term water demand prediction. The proposed method consists of four main parts: the variational mode decomposition method, the golden jackal optimization algorithm, the multihead attention mechanism, and the bidirectional gated recurrent unit (BiGRU) model. Furthermore, a seq2seq strategy was adopted for multistep prediction to avoid the error accumulation problem. Hourly water demand data collected from a real-world water distribution system were applied to investigate the potential of the proposed method. The results show that the proposed method can yield remarkably accurate and stable forecasts in single-step prediction (i.e., the mean absolute percentage error (MAPE) reaches 0.45%, and the root mean squared error (RMSE) is 25 m3/h). Moreover, the proposed method still achieves credible performance in 24-step prediction (i.e., the MAPE reaches 2.12%, and the RMSE is 126 m3/h). In general, for both single-step prediction and multistep prediction, the proposed method consistently outperforms other BiGRU-based methods. These findings suggest that the proposed method can provide a reliable alternative for short-term water demand prediction.

PatchesNet: PatchTST-based multi-scale network security situation prediction

Internet of Things (IoT) technology increases connectivity between devices; however, IoT device networks are less secure, making them highly susceptible to attack. Most of the network security defense techniques are passive defense with inevitable delays. For instance, in the case of intrusion detection systems, during the time interval between detecting an attack and implementing defenses, the attacker may inflict damage on the target. Therefore, the adoption of active defense techniques becomes particularly important. Network Security Situation Prediction (NSSP) is an active defense technology capable of predicting future states of network security. Currently, most methods in the NSSP field focus on single-step prediction, while research related to multi-step prediction is scarce. Hence, this paper presents an NSSP method based on the PatchTST framework(PatchesNet), which exhibits good multi-step predictive performance. The method uses Variational Mode Decomposition (VMD) to deal with sequence instability. Multi-scale Patch delivery (MSP), Multi-scale Second Decomposition (MSD), and Period Embedding (PeE) are used to obtain more sequence features. Correlation Fusion (CF) can reduce the noise effect on prediction. Experiments are conducted on three datasets. The results demonstrate that PatchesNet exhibits favorable performance, relatively reducing the MSE and MAE by 39.243% and 26.033%, respectively, compared to the previous state-of-the-art method.

Research on the Fiber-to-the-Room Network Traffic Prediction Method Based on Crested Porcupine Optimizer Optimization

In order to solve the problem of traffic burst due to the increase in access points and user movement in an FTTR network, as well as to meet the demand for a high-performance network, it is necessary to rationally allocate network resources, and accurate traffic prediction is very important for dynamic bandwidth allocation in such a network. Therefore, this paper introduces a novel traffic prediction model, named CPO-BiTCN-BiLSTM-SA, which integrates the Crested Porcupine Optimizer (CPO), bidirectional temporal convolution (BiTCN), and bidirectional long short-term memory (BiLSTM) networks. BiTCN extends the traditional TCN by incorporating bidirectional data information, while BiLSTM enhances the network’s capability to learn from long sequences. Moreover, self-attention (SA) mechanisms are utilized to emphasize the crucial segments in the data. Subsequently, the BiTCN-BiLSTM-SA model is optimized by CPO to obtain the best network hyperparameters, and model training prediction is performed to achieve multi-step predictions based on single-step prediction. To evaluate the model’s generalization ability, two distinct datasets are employed for traffic prediction. Experimental findings demonstrate that the proposed model surpasses existing models in terms of the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). In comparison with the traditional XGBoost model, the proposed model has an average reduction of 29.50%, 25.43%, and 25.00% in RMSE, MAE, and MAPE, respectively, with a 6.70% improvement in R2.

Growing strings in a chemical reaction space for searching retrosynthesis pathways

Machine learning algorithms have shown great accuracy in predicting chemical reaction outcomes and retrosyntheses. However, designing synthesis pathways remains challenging for existing machine learning models which are trained for single-step prediction. In this manuscript, we propose to recast the retrosynthesis problem as a string optimization problem in a data-driven fingerprint space, leveraging the similarity between chemical reactions and embedding vectors. Based on this premise, multi-step complex synthesis can be conceptualized as sequences that link multidimensional vectors (fingerprints) representing individual chemical reaction steps. We extracted an extensive corpus of chemical synthesis from patents and converted them into multidimensional strings. While optimizing the retrosynthetic path, we use the Euclidean metric to minimize the distance between the expanded trajectory of the growing retrosynthesis string and the corpus of extracted strings. By doing so, we promote the assembly of synthetic pathways that, in the chemical reaction space, will be more similar to existing retrosyntheses, thereby inheriting the strategic guidelines designed by human experts. We integrated this approach into the RXN platform (https://rxn.res.ibm.com/) and present the method’s application to complex synthesis as well as its ability to produce better synthetic strategies than current methodologies.

Optimizing energy storage plant discrete system dynamics analysis with graph convolutional networks

Addressing the challenges of suboptimal model performance and excessive parameters and operations in the optimization of energy storage power plants utilizing Graph Convolutional Network (GCN), this paper introduces a novel approach – the packet-switched graph convolutional network. Initially, a GCN extreme learning machine is established. Drawing inspiration from this solid foundation, we have innovatively crafted a group exchange graph convolution module. This module leverages group graph convolution techniques to amalgamate unique node feature information, tailored to diverse topology graph matrices based on various groupings. This innovative approach ensures that information flows freely and effectively among distinct groupings. Furthermore, we have designed a cutting-edge timing depth separation convolution module, comprising two innovative components. The first component introduces timing depth separation convolution, revolutionizing the original timing convolution module. The second component, the packet-switching graph convolutional network, revolutionizes the time sequence depth separation convolution process. It achieves this by employing 1 × 1 convolutional layers between different feature fusion packets, enabling seamless information exchange between distinct packets. Experimental results demonstrate the efficacy of the proposed model, with root mean square error (RMSE) metrics and root mean square error (MAE) metrics for single-step prediction reaching 46.08 and 26.22 at 60 min, respectively. In multi-step testing, the proposed model exhibits a 14.71 % reduction in RMSE error at the 15-min scale and a 9.29 % reduction at the 60-min scale compared to the benchmark model. This performance improvement enhances the operational efficiency and reliability of the energy storage plant, particularly under dynamic changes in the time series.

Transfer learning integrating similarity analysis for short-term and long-term building energy consumption prediction

Currently, building energy consumption prediction models are usually based on a large amount of historical operational data in high demands of building operating hours and monitoring systems. However, many buildings may lack operational data due to relatively limited monitoring systems, causing the failure to use data-driven methods to characterize the energy profile. In this context, transfer learning is a promising method to establish the knowledge transfer between many high-quality building operation datasets and a small amount of target building data, and to help predict energy consumption in the target building. This paper studies the possibility of employing transfer learning to achieve both short and long-term building energy consumption prediction. Firstly, a similarity analysis method, based on variable modal decomposition and dynamic time warping, is proposed for identifying the source buildings with the most similar energy features to the target building. Then, transfer learning for long-term prediction air-conditioning energy consumption is developed with weather parameters generated by the Morphing method as inputs. For the short-term single-step prediction, the proposed model CV-RMSE improves 81.3% (AEC) and 77.4% (EEC), respectively, compared to the prediction model that does not implement the transfer learning strategy and directly uses the target BEC data. As for the short-term multi-step prediction, the proposed model CV-RMSE improves 62.0% (AEC) and 65.5% (EEC), respectively. For the long-term prediction, the average CV-RMSE for the whole year is 6.62% and 11.15% for the proposed and directly target domain-based model, respectively. The proposed method explores the practicality of transfer learning in building energy forecasting, contributing to the use of existing building operation data for energy management at different timespan.

The Carbon Emission Accounting and Prediction of the Power Generation Side based on LSTM in Jilin Province

In the context of global warming and the dramatic increase in greenhouse gas emissions, the power industry is the largest source of carbon emissions. Adjusting and optimizing the carbon dioxide emissions from the power industry will help China achieve its “dual carbon” goals and is of great significance for mitigating global carbon dioxide emissions. This paper takes six power plants in Jilin Province as the research objects, and firstly accounts for the carbon emission production data between January 2020 and December 2023 according to the "Accounting Methods and Reporting Guidelines for Greenhouse Gas Emissions from Enterprises - Power Generation Facilities". Then, LSTM was used to establish carbon emission prediction models for six different power plants in Jilin Province, and the analysis of each model showed that that the single-step prediction RMSEs are all less than one, with higher prediction accuracy, but only can used in short-term prediction, the multi-step prediction RMSEs are bigger than one, with lower prediction accuracy, but can used in long-term carbon emission trend prediction can be achieved. The carbon emission trend prediction of six power plants in Jilin province between January 2024 and August 2031 confirms that the carbon emissions of power plants will be affected by seasons and shown cyclical changes. Finally, reasonable policy recommendations are provided for the successful realisation of the "double carbon" target for electricity in Jilin Province.

An Intelligent Early Warning System for Harmful Algal Blooms: Harnessing the Power of Big Data and Deep Learning.

Harmful algal blooms (HABs) pose a significant ecological threat and economic detriment to freshwater environments. In order to develop an intelligent early warning system for HABs, big data and deep learning models were harnessed in this study. Data collection was achieved utilizing the vertical aquatic monitoring system (VAMS). Subsequently, the analysis and stratification of the vertical aquatic layer were conducted employing the "DeepDPM-Spectral Clustering" method. This approach drastically reduced the number of predictive models and enhanced the adaptability of the system. The Bloomformer-2 model was developed to conduct both single-step and multistep predictions of Chl-a, integrating the " Alert Level Framework" issued by the World Health Organization to accomplish early warning for HABs. The case study conducted in Taihu Lake revealed that during the winter of 2018, the water column could be partitioned into four clusters (Groups W1-W4), while in the summer of 2019, the water column could be partitioned into five clusters (Groups S1-S5). Moreover, in a subsequent predictive task, Bloomformer-2 exhibited superiority in performance across all clusters for both the winter of 2018 and the summer of 2019 (MAE: 0.175-0.394, MSE: 0.042-0.305, and MAPE: 0.228-2.279 for single-step prediction; MAE: 0.184-0.505, MSE: 0.101-0.378, and MAPE: 0.243-4.011 for multistep prediction). The prediction for the 3 days indicated that Group W1 was in a Level I alert state at all times. Conversely, Group S1 was mainly under an Level I alert, with seven specific time points escalating to a Level II alert. Furthermore, the end-to-end architecture of this system, coupled with the automation of its various processes, minimized human intervention, endowing it with intelligent characteristics. This research highlights the transformative potential of integrating big data and artificial intelligence in environmental management and emphasizes the importance of model interpretability in machine learning applications.

A Deep U-Net-ConvLSTM Framework with Hydrodynamic Model for Basin-Scale Hydrodynamic Prediction

Traditional hydrodynamic models face the significant challenge of balancing the demands of long prediction spans and precise boundary conditions, large computational areas, and low computational costs when attempting to rapidly and accurately predict the nonlinear spatial and temporal characteristics of fluids at the basin scale. To tackle this obstacle, this study constructed a novel deep learning framework with a hydrodynamic model for the rapid spatiotemporal prediction of hydrodynamics at the basin scale, named U-Net-ConvLSTM. A validated high-fidelity hydrodynamic mechanistic model was utilized to build a 20-year hydrodynamic indicator dataset of the middle and lower reaches of the Han River for the training and validation of U-Net-ConvLSTM. The findings indicate that the R2 value of the model surpassed 0.99 when comparing the single-step prediction results with the target values. Additionally, the required computing time fell by 62.08% compared with the hydrodynamic model. The ablation tests demonstrate that the U-Net-ConvLSTM framework outperforms other frameworks in terms of accuracy for basin-scale hydrodynamic prediction. In the multi-step-ahead prediction scenarios, the prediction interval increased from 1 day to 5 days, while consistently maintaining an R2 value above 0.7, which demonstrates the effectiveness of the model in the missing boundary conditions scenario. In summary, the U-Net-ConvLSTM framework is capable of making precise spatiotemporal predictions in hydrodynamics, which may be considered a high-performance computational solution for predicting hydrodynamics at the basin scale.