Deep Learning Prediction Research Articles

Regional aerosol forecasts based on deep learning and numerical weather prediction

Atmospheric chemistry transport models have been extensively applied in aerosol forecasts over recent decades, whereas they are facing challenges from uncertainties in emission rates, meteorological data, and over-simplified chemical parameterizations. Here, we developed a spatial-temporal deep learning framework, named PPN (Pollution-Predicting Net for PM2.5), to accurately and efficiently predict regional PM2.5 concentrations. It has an encoder-decoder architecture and combines the preceding PM2.5 observations and numerical weather prediction. Besides, the model proposes a weighted loss function to promote the forecasting performance in extreme events. We applied the proposed model to forecast 3-day PM2.5 concentrations over the Beijing-Tianjin-Hebei region in China on a three-hour-by-three-hour basis. Overall, the model showed good performance with R2 and RMSE values of 0.7 and 17.7 μg m−3, respectively. It could capture the high PM2.5 concentration in the south and relatively low concentration in the north and exhibit better performance within the next 24 h. The use of the weighted loss function decreased the level of “high values underestimation, low values overestimation”, while incorporating the preceding PM2.5 observations into the encoder phase improved the predictive accuracy within 24 h. We also compared the model result with that from a state-of-the-art numerical model (WRF-Chem with pollutant data assimilation). The temporal R2 and RMSE from the WRF-Chem were 0.30−0.77 and 19−45 μg m−3 while those from the PPN model were 0.42−0.84 and 15−42 μg m−3. The proposed model shows powerful capacity in aerosol forecasts and provides an efficient and accurate tool for early warning and management of regional pollution events.

Spatiotemporal Prediction of Ionospheric Total Electron Content Based on ED-ConvLSTM

Total electron content (TEC) is a vital parameter for describing the state of the ionosphere, and precise prediction of TEC is of great significance for improving the accuracy of the Global Navigation Satellite System (GNSS). At present, most deep learning prediction models just consider TEC temporal variation, while ignoring the impact of spatial location. In this paper, we propose a TEC prediction model, ED-ConvLSTM, which combines convolutional neural networks with recurrent neural networks to simultaneously consider spatiotemporal features. Our ED-ConvLSTM model is built based on the encoder-decoder architecture, which includes two modules: encoder module and decoder module. Each module is composed of ConvLSTM cells. The encoder module is used to extract the spatiotemporal features from TEC maps, while the decoder module converts spatiotemporal features into predicted TEC maps. We compared the predictive performance of our model with two traditional time series models: LSTM, GRU, a spatiotemporal mode1 ConvGRU, and the TEC daily forecast product C1PG provided by CODE on a total of 135 grid points in East Asia (10°–45°N, 90°–130°E). The experimental results show that the prediction error indicators MAE, RMSE, MAPE, and prediction similarity index SSIM of our model are superior to those of the comparison models in high, normal, and low solar activity years. The paper also analyzed the predictive performance of each model monthly. The experimental results indicate that the predictive performance of each model is influenced by the monthly mean of TEC. The ED-ConvLSTM model proposed in this paper is the least affected and the most stable by the monthly mean of TEC. Additionally, the paper compared the predictive performance of each model during two magnetic storm periods when TEC changes sharply. The results indicate that our ED-ConvLSTM model is least affected during magnetic storms and its predictive performance is superior to those of the comparative models. This paper provides a more stable and high-performance TEC spatiotemporal prediction model.

Deep learning prediction intervals based on selective joint supervision

Deep learning prediction intervals based on selective joint supervision

Dual-dimension Time-GGAN data augmentation method for improving the performance of deep learning models for PV power forecasting

Solar energy has intermittency and volatility, which poses a serious challenge to the stable and safe operation of the power grid. Accurate photovoltaic power prediction is one of the key technologies to solve this problem. Due to the larger volume of data collected by photovoltaic systems, deep learning methods are more in-depth in the field of photovoltaic power prediction. However, photovoltaic power generation data also has more diversity and lower value density, which brings more severe challenges to the feature extraction ability of prediction models. Photovoltaic power prediction methods with deep data mining capabilities are urgently needed to address these issues. This paper proposes a dual-dimensional time series adversarial neural network based on gated recurrent neural network (Dd_Time_GGAN) to enhance low-value-density photovoltaic data by two dimensions (time dimension and feature dimension) and obtain high-value-density feature data. by replacing the generator and optimizer to improve the performance of network-enhanced data, the training time of the network was reduced by 26% by adjusting the iteration times of each stage of the network. Then, combining the two mainstream deep neural networks LSTM and CNN models in the current photovoltaic forecast field to achieve ultra-short-term photovoltaic power prediction. The experimental data is two-year operation data of a photovoltaic station in China. The experimental results show that compared with only using raw sample data, using enhanced data proposed by the data augmentation method combined with raw sample data as input, both types of deep learning prediction models have higher prediction accuracy. The accuracy of the LSTM prediction model improved by 3.1%, and the CNN prediction model improved by 2.3%. It proves that through data enhancement, the prediction accuracy of the photovoltaic power prediction model can be improved.

Developing Spatial and Temporal Continuous Fractional Vegetation Cover Based on Landsat and Sentinel-2 Data with a Deep Learning Approach

Fractional vegetation cover (FVC) has a significant role in indicating changes in ecosystems and is useful for simulating growth processes and modeling land surfaces. The fine-resolution FVC products represent detailed vegetation cover information within fine grids. However, the long revisit cycle of satellites with fine-resolution sensors and cloud contamination has resulted in poor spatial and temporal continuity. In this study, we propose to derive a spatially and temporally continuous FVC dataset by comparing multiple methods, including the data-fusion method (STARFM), curve-fitting reconstruction (S-G filtering), and deep learning prediction (Bi-LSTM). By combining Landsat and Sentinel-2 data, the integrated FVC was used to construct the initial input of fine-resolution FVC with gaps. The results showed that the FVC of gaps were estimated and time-series FVC was reconstructed. The Bi-LSTM method was the most effective and achieved the highest accuracy (R2 = 0.857), followed by the data-fusion method (R2 = 0.709) and curve-fitting method (R2 = 0.705), and the optimal time step was 3. The inclusion of relevant variables in the Bi-LSTM model, including LAI, albedo, and FAPAR derived from coarse-resolution products, further reduced the RMSE from 5.022 to 2.797. By applying the optimized Bi-LSTM model to Hubei Province, a time series 30 m FVC dataset was generated, characterized by a spatial and temporal continuity. In terms of the major vegetation types in Hubei (e.g., evergreen and deciduous forests, grass, and cropland), the seasonal trends as well as the spatial details were captured by the reconstructed 30 m FVC. It was concluded that the proposed method was applicable to reconstruct the time-series FVC over a large spatial scale, and the produced fine-resolution dataset can support the data needed by many Earth system science studies.

An Intelligent Approach to Short-Term Wind Power Prediction Using Deep Neural Networks

Abstract In this paper, an intelligent approach to the Short-Term Wind Power Prediction (STWPP) problem is considered, with the use of various types of Deep Neural Networks (DNNs). The impact of the prediction time horizon length on accuracy, and the influence of temperature on prediction effectiveness have been analyzed. Three types of DNNs have been implemented and tested, including: CNN (Convolutional Neural Networks), GRU (Gated Recurrent Unit), and H-MLP (Hierarchical Multilayer Perceptron). The DNN architectures are part of the Deep Learning Prediction (DLP) framework that is applied in the Deep Learning Power Prediction System (DLPPS). The system is trained based on data that comes from a real wind farm. This is significant because the prediction results strongly depend on weather conditions in specific locations. The results obtained from the proposed system, for the real data, are presented and compared. The best result has been achieved for the GRU network. The key advantage of the system is a high effectiveness prediction using a minimal subset of parameters. The prediction of wind power in wind farms is very important as wind power capacity has shown a rapid increase, and has become a promising source of renewable energies.

Improvement of deep learning prediction model in patient-specific QA for VMAT with MLC leaf position map and patient's dose distribution.

Deep learning-based virtual patient-specific quality assurance (QA) is a novel technique that enables patient QA without measurement. However, this method could be improved by further evaluating the optimal data to be used as input. Therefore, a deep learning-based model that uses multileaf collimator (MLC) information per control point and dose distribution in patient's CT as inputs was developed. Overall, 96 volumetric-modulated arc therapy plans generated for prostate cancer treatment were used. We developed a model (Model 1) that can predict measurement-based gamma passing rate (GPR) for a treatment plan using data stored as a map reflecting the MLC leaf position at each control point (MLPM) and data of the dose distribution in patient's CT as inputs. The evaluation of the model was based on the mean absolute error (MAE) and Pearson's correlation coefficient (r) between the measured and predicted GPR. For comparison, we also analyzed models trained with the dose distribution in patient's CT alone (Model 2) and with dose distributions recalculated on a virtual phantom CT (Model 3). At the 2%/2mm criterion, MAE[%] and r for Model 1, Model 2, and Model 3 were 2.32%±0.43% and 0.54±0.03, 2.70%±0.26%, and 0.32±0.08, and 2.96%±0.23% and 0.24±0.22, respectively; at the 3%/3mm criterion, these values were 1.25%±0.05% and 0.36±0.18, 1.57%±0.35% and 0.19±0.20, and 1.39%±0.32% and 0.17±0.22, respectively. This result showed that Model 1 exhibited the lowest MAE and highest r at both criteria of 2%/2mm and 3%3mm. These findings showed that a model that combines the MLPM and dose distribution in patient's CT exhibited a better GPR prediction performance compared with the other two studied models.

THU-491 - Deep learning prediction modeling of major adverse cardiovascular events following liver transplantation

THU-491 - Deep learning prediction modeling of major adverse cardiovascular events following liver transplantation

Electromagnetic Interference (EMI) Elimination via Active Sensing and Deep Learning Prediction for RF Shielding-free MRI.

At present, MRI scans are typically performed inside fully enclosed RF shielding rooms, posing stringent installation requirement and patient discomfort. We aim to eliminate electromagnetic interference (EMI) for MRI with no or incomplete RF shielding. In this study, a method of active sensing and deep learning EMI prediction is presented to model, predict and remove EMI signal components from acquired MRI signals. Specifically, during each MRI scan, separate EMI sensing coils placed in various locations are utilized to simultaneously sample external and internal EMI signals within two windows (for both conventional MRI signal acquisition and EMI characterization acquisition). A CNN model is trained using the EMI characterization data to relate EMI signals detected by EMI sensing coils to EMI signals in MRI receive coil. This model is then used to retrospectively predict and remove EMI signal components detected by MRI receive coil during the MRI signal acquisition window. This strategy was implemented on a low-cost ultra-low-field 0.055 T permanent magnet MRI scanner without RF shielding. It produced final image signal-to-noise ratios that were comparable to those obtained using a fully enclosed RF shielding cage, and outperformed existing analytical EMI elimination methods (i.e., transfer function and EDITER methods). A preliminary experiment also demonstrated its applicability on a 1.5 T superconducting magnet MRI scanner with incomplete RF shielding. Altogether, the results demonstrated that the proposed method was highly effective in predicting and removing various EMI signals from both external environments and internal scanner electronics at both 0.055 T (2.3 MHz) and 1.5 T (63.9 MHz). The proposed strategy enables shielding-free MRI. The concept is relatively simple and potentially applicable to other RF signal detection scenarios in presence of external or/and internal EMI.

Energy Management Strategy Based on Deep Reinforcement Learning and Speed Prediction for Power‐Split Hybrid Electric Vehicle with Multidimensional Continuous Control

An efficient energy management strategy (EMS) is significant to improve the economy of hybrid electric vehicles (HEVs). Herein, a power‐split HEV model is built and validated against test results, and then the EMS is proposed for this model based on vehicle speed prediction and deep reinforcement learning (DRL) algorithms. The rule‐based local controller and global optimal empirical knowledge are introduced to enhance the convergence speed. It is shown in the results that the twin delayed deep deterministic policy gradient algorithm (TD3) achieves more satisfactory performance on converge speed and energy efficiency. The networks of the DRL algorithm with continuous control update more robustly during iterations, in contrast to the discrete ones. Although the power‐split HEV with lower control dimension can reduce the learning burden for DRL EMS; however, the multidimensional control space shows greater optimization potential. As a result, the equivalent fuel consumption of TD3‐based EMS with multidimensional continuous control differences from the global optimal algorithm only by 4.92%. Herein, it is demonstrated in the results that long short‐term memory recurrent neural network (LSTM RNN) performs better for vehicle speed prediction than classical RNN and BP neural network, and the predictive vehicle speed feature helps improve fuel economy by 0.55%.

CT-based deep multi-label learning prediction model for outcome in patients with oropharyngeal squamous cell carcinoma.

Personalized treatment is increasingly required for oropharyngeal squamous cell carcinoma (OPSCC) patients due to emerging new cancer subtypes and treatment options. Outcome prediction model can help identify low or high-risk patients who may be suitable to receive de-escalation or intensified treatment approaches. To develop a deep learning (DL)-based model for predicting multiple and associated efficacy endpoints in OPSCC patients based on computed tomography (CT). Two patient cohorts were used in this study: a development cohort consisting of 524 OPSCC patients (70% for training and 30% for independent testing) and an external test cohort of 396 patients. Pre-treatment CT-scans with the gross primary tumor volume contours (GTVt) and clinical parameters were available to predict endpoints, including 2-year local control (LC), regional control (RC), locoregional control (LRC), distant metastasis-free survival (DMFS), disease-specific survival (DSS), overall survival (OS), and disease-free survival (DFS). We proposed DL outcome prediction models with the multi-label learning (MLL) strategy that integrates the associations of different endpoints based on clinical factors and CT-scans. The multi-label learning models outperformed the models that were developed based on a single endpoint for all endpoints especially with high AUCs≥0.80 for 2-year RC, DMFS, DSS, OS, and DFS in the internal independent test set and for all endpoints except 2-year LRC in the external test set. Furthermore, with the models developed, patients could be stratified into high and low-risk groups that were significantly different for all endpoints in the internal test set and for all endpoints except DMFS in the external test set. MLL models demonstrated better discriminative ability for all 2-year efficacy endpoints than single outcome models in the internal test and for all endpoints except LRC in the external set.

Short-term Bus Passenger Flow Forecast Based on CNN-BiLSTM

Effective prediction of urban bus passenger flow is critical for improving urban bus operation efficiency and optimizing the bus network. However, there are some issues with predicting urban bus passenger flow at the moment, such as lack of single eigenvalue consideration and insufficient research depth. In order to improve the short-term prediction accuracy of urban bus passenger flow, this paper proposed a deep learning prediction model that is based on CNN-BiLSTM. Based on historical data of urban bus passenger flow, this paper analyzes the dependence of bus credit card data, clusters the travel feature of different groups of people, and analyzes the dependence of bus passengers. Simultaneously, external factors of passenger flow, such as rainfall, weather condition, traffic flow state, and date, are introduced to build the bus passenger flow prediction feature matrix, and the correlation analysis of the characteristic matrix structure is performed to optimize the matrix structure. Finally, the optimized passenger flow characteristic matrix is fed into the CNN-BiLSTM deep learning model for prediction, and the results are compared to the LSTM, CNN and CNN-LSTM models. The results shown that the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the the CNN-BiLSTM deep learning model are lower than those of other models, and the prediction accuracy is the highest. Meanwhile, this method has a good generalization effect and can improve deep learning prediction accuracy.

Common vulnerability scoring system prediction based on open source intelligence information sources

The number of newly published vulnerabilities is constantly increasing. Until now, the information available when a new vulnerability is published is manually assessed by experts using a Common Vulnerability Scoring System (CVSS) vector and score. This assessment is time consuming and requires expertise. Various works already try to predict CVSS vectors or scores using machine learning based on the textual descriptions of the vulnerability to enable faster assessment. However, for this purpose, previous works only use the texts available in databases such as National Vulnerability Database. With this work, the publicly available web pages referenced in the National Vulnerability Database are analyzed and made available as sources of texts through web scraping. A Deep Learning based method for predicting the CVSS vector is implemented and evaluated. The present work provides a classification of the National Vulnerability Database’s reference texts based on the suitability and crawlability of their texts. While we identified the overall influence of the additional texts is negligible, we outperformed the state-of-the-art with our Deep Learning prediction models.

Grad-MobileNet: A Gradient-Based Unsupervised Learning Method for Laser Welding Surface Defect Classification.

Deep learning technology has advanced rapidly and has started to be applied for the detection of welding defects. In the manufacturing process of power batteries for new energy vehicles, welding defects may occur due to the high directivity, convergence, and penetration of the laser beam. The accuracy of deep learning prediction relies heavily on big data, but balanced big data of welding defects is hard to acquire at the battery production site. In this paper, the authors construct a dataset named RIAM, which consists of images captured from an industrial environment for laser welding of power battery modules. RIAM contains four types of images: Normality, Lack of fusion, Surface porosity, and Scaled surface. The characteristics of RIAM are carefully considered in the application scenarios. Moreover, this paper proposes a gradient-based unsupervised model named Grad-MobileNet, which can be trained with only a few normal images and can extract the feature gradients of the input images. Welding defects can then be classified by the gradient distribution. This model is based on MobileNetV3, which is a lightweight convolutional neural network (CNN), and achieves 99% accuracy, which is higher than the accuracy expected from supervised learning.

Patient selection for proton therapy using Normal Tissue Complication Probability with deep learning dose prediction for oropharyngeal cancer.

In cancer care, determining the most beneficial treatment technique is a key decision affecting the patient's survival and quality of life. Patient selection for proton therapy (PT) over conventional radiotherapy (XT) currently entails comparing manually generated treatment plans, which requires time andexpertise. We developed an automatic and fast tool, AI-PROTIPP (Artificial Intelligence Predictive Radiation Oncology Treatment Indication to Photons/Protons), that assesses quantitatively the benefits of each therapeutic option. Our method uses deep learning (DL) models to directly predict the dose distributions for a given patient for both XT and PT. By using models that estimate the Normal Tissue Complication Probability (NTCP), namely the likelihood of side effects to occur for a specific patient, AI-PROTIPP can propose a treatment selection quickly andautomatically. A database of 60 patients presenting oropharyngeal cancer, obtained from the Cliniques Universitaires Saint Luc in Belgium, was used in this study. For every patient, a PT plan and an XT plan were generated. The dose distributions were used to train the two dose DL prediction models (one for each modality). The model is based on U-Net architecture, a type of convolutional neural network currently considered as the state of the art for dose prediction models. A NTCP protocol used in the Dutch model-based approach, including grades II and III xerostomia and grades II and III dysphagia, was later applied in order to perform automatic treatment selection for each patient. The networks were trained using a nested cross-validation approach with 11-folds. We set aside three patients in an outer set and each fold consists of 47 patients in training, five in validation and five for testing. This method allowed us to assess our method on 55 patients (five patients per test times the number of folds). The treatment selection based on the DL-predicted doses reached an accuracy of 87.4% for the threshold parameters set by the Health Council of the Netherlands. The selected treatment is directly linked with these threshold parameters as they express the minimal gain brought by the PT treatment for a patient to be indicated to PT. To validate the performance of AI-PROTIPP in other conditions, we modulated these thresholds, and the accuracy was above 81% for all the considered cases. The difference in average cumulative NTCP per patient of predicted and clinical dose distributions is very similar (less than 1% difference). AI-PROTIPP shows that using DL dose prediction in combination with NTCP models to select PT for patients is feasible and can help to save time by avoiding the generation of treatment plans only used for the comparison. Moreover, DL models are transferable, allowing, in the future, experience to be shared with centers that would not have PT planningexpertise.

On the Early and Affordable Diagnosis of Joint Pathologies Using Acoustic Emissions, Deep Learning Decompositions and Prediction Machines.

The condition of a joint in a human being is prone to wear and several pathologies, particularly in the elderly and athletes. Current means towards assessing the overall condition of a joint to assess for a pathology involve using tools such as X-ray and magnetic resonance imaging, to name a couple. These expensive methods are of limited availability in resource-constrained environments and pose the risk of radiation exposure to the patient. The prospect of acoustic emissions (AEs) presents a modality that can monitor the joints' conditions passively by recording the high-frequency stress waves emitted during their motion. One of the main challenges associated with this sensing method is decoding and linking acquired AE signals to their source event. In this paper, we investigate AEs' use to identify five kinds of joint-wear pathologies using a contrast of expert-based handcrafted features and unsupervised feature learning via deep wavelet decomposition (DWS) alongside 12 machine learning models. The results showed an average classification accuracy of 90 ± 7.16% and 97 ± 3.77% for the handcrafted and DWS-based features, implying good prediction accuracies across the various devised approaches. Subsequent work will involve the potential application of regressions towards estimating the associated stage and extent of a wear condition where present, which can form part of an online system for the condition monitoring of joints in human beings.

Research on refined prediction of coastal wind power based on dynamic downscale and deep learning prediction

With the expansion of the installed capacity of offshore wind power, its impact on the safe operation of the power grid is increasingly obvious, and improving the prediction accuracy of offshore wind power power has become a key problem to be solved in the industry. The improvement of prediction accuracy, on the one hand, can facilitate the operation control of power grid, on the other hand, can also significantly improve the economic benefits of wind power plants. Based on power downscaling and deep learning methods, this paper carries out localized short-term and ultra-short-term power prediction modeling and optimization technologies for coastal offshore wind power. Comprehensive statistical prediction method, artificial intelligence prediction method, multi-model integration method and other combined prediction model, multi-source fusion information, to realize the refined prediction of Zhejiang offshore wind power, and improve the overall prediction accuracy. The research method is generally applicable and can guide the output prediction of the same type of offshore wind power.

PO-2123 Deep Learning Prediction of 2-year Survival in Lung Cancer Patients Undergoing Radical Radiotherapy

PO-2123 Deep Learning Prediction of 2-year Survival in Lung Cancer Patients Undergoing Radical Radiotherapy

DLC-ac4C: A Prediction Model for N4-acetylcytidine Sites in Human mRNA Based on DenseNet and Bidirectional LSTM Methods.

N4 acetylcytidine (ac4C) is a highly conserved nucleoside modification that is essential for the regulation of immune functions in organisms. Currently, the identification of ac4C is primarily achieved using biological methods, which can be time-consuming and labor-intensive. In contrast, accurate identification of ac4C by computational methods has become a more effective method for classification and prediction. To the best of our knowledge, although there are several computational methods for ac4C locus prediction, the performance of the models they constructed is poor, and the network structure they used is relatively simple and suffers from the disadvantage of network degradation. This study aims to improve these limitations by proposing a predictive model based on integrated deep learning to better help identify ac4C sites. In this study, we propose a new integrated deep learning prediction framework, DLC-ac4C. First, we encode RNA sequences based on three feature encoding schemes, namely C2 encoding, nucleotide chemical property (NCP) encoding, and nucleotide density (ND) encoding. Second, one-dimensional convolutional layers and densely connected convolutional networks (DenseNet) are used to learn local features, and bi-directional long short-term memory networks (Bi-LSTM) are used to learn global features. Third, a channel attention mechanism is introduced to determine the importance of sequence characteristics. Finally, a homomorphic integration strategy is used to limit the generalization error of the model, which further improves the performance of the model. The DLC-ac4C model performed well in terms of sensitivity (Sn), specificity (Sp), accuracy (Acc), Mathews correlation coefficient (MCC), and area under the curve (AUC) for the independent test data with 86.23%, 79.71%, 82.97%, 66.08%, and 90.42%, respectively, which was significantly better than the prediction accuracy of the existing methods. Our model not only combines DenseNet and Bi-LSTM, but also uses the channel attention mechanism to better capture hidden information features from a sequence perspective, and can identify ac4C sites more effectively.

Evolution of neural network to deep learning in prediction of air, water pollution and its Indian context.

The scenario of developed and developing countries nowadays is disturbed due to modern living style which affects environment, wildlife and natural habitat. Environmental quality has become or is a subject of major concern as it is responsible for health hazard of mankind and animals. Measurements and prediction of hazardous parameters in different fields of environment is a recent research topic for safety and betterment of people as well as nature. Pollution in nature is an after-effect of civilization. To combat the damage already happened, some processes should be evolved for measurement and prediction of pollution in various fields. Researchers of all over the world are active to find out ways of predicting such hazard. In this paper, application of neural network and deep learningalgorithms is chosen for air pollution and water pollution cases. The purpose of this review isto reveal how family of neural network algorithms has applied on these two pollution parameters. In this paper, importance is given on algorithm, and datasets used for air and water pollution as well as the predicted parameters have also been noted for ease of future development. One major concern of this paper is Indian context of air and water pollution research, and the research potential presents in this area using Indian dataset. Another aspect for including both air and water pollutions in one review paper is to generate an idea of artificial neural network and deep learning techniques which can be cross applicable for future purpose.