Causal Convolutional Neural Network Research Articles

Novel spatio-temporal attention causal convolutional neural network for multi-site PM2.5 prediction

Multi-site PM2.5 prediction has emerged as a crucial approach, given that the accuracy of prediction models based solely on data from a single monitoring station may be constrained. However, existing multi-site PM2.5 prediction methods predominantly rely on recurrent networks for extracting temporal dependencies and overlook the domain knowledge related to air quality pollutant dispersion. This study aims to explore whether a superior prediction architecture exists that not only approximates the prediction performance of recurrent networks through feedforward networks but also integrates domain knowledge of PM2.5. Consequently, we propose a novel spatio-temporal attention causal convolutional neural network (Causal-STAN) architecture for predicting PM2.5 concentrations at multiple sites in the Yangtze River Delta region of China. Causal-STAN comprises two components: a multi-site spatio-temporal feature integration module, which identifies temporal local correlation trends and spatial correlations in the spatio-temporal data, and extracts inter-site PM2.5 concentrations from the directional residual block to delineate directional features of PM2.5 concentration dispersion between sites; and a temporal causal attention convolutional network that captures the internal correlation information and long-term dependencies in the time series. Causal-STAN was evaluated using one-year data from 247 sites in mainland China. Compared to six state-of-the-art baseline models, Causal-STAN achieves optimal performance in 6-hour future predictions, surpassing the recurrent network model and reducing the prediction error by 8%–10%.

Daily PM2.5 concentration prediction based on variational modal decomposition and deep learning for multi-site temporal and spatial fusion of meteorological factors.

Air pollution, particularly PM2.5, has long been a critical concern for the atmospheric environment. Accurately predicting daily PM2.5 concentrations is crucial for both environmental protection and public health. This study introduces a new hybrid model within the "Decomposition-Prediction-Integration" (DPI) framework, which combines variational modal decomposition (VMD), causal convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and attention mechanism (AM), named as VCBA, for spatio-temporal fusion of multi-site data to forecast daily PM2.5 concentrations in a city. The approach involves integrating air quality data from the target site with data from neighboring sites, applying mathematical techniques for dimensionality reduction, decomposing PM2.5 concentration data using VMD, and utilizing Causal CNN and BiLSTM models with an attention mechanism to enhance performance. The final prediction results are obtained through linear aggregation. Experimental results demonstrate that the VCBA model performs exceptionally well in predicting daily PM2.5 concentrations at various stations in Taiyuan City, Shanxi Province, China. Evaluation metrics such as RMSE, MAE, and R2 are reported as 2.556, 1.998, and 0.973, respectively. Compared to traditional methods, this approach offers higher prediction accuracy and stronger spatio-temporal modeling capabilities, providing an effective solution for accurate PM2.5 daily concentration prediction.

Self-testing Algorithm of Metro One-button Switch Station Equipment Based on a Neural Network

According to the operation and management requirements of subway stations, equipment managed by the station, such as lighting, escalators, broadcasting, etc., must be operated with one-button switch according to a certain process every day. The first step in one-button switch station operation is to perform routine self-testing on these devices. Traditional self-testing operations mainly rely on manual handling by operators, which greatly affects their work efficiency. To optimize the operational processes, it is necessary to establish intelligent self-testing algorithms. An AC-CNN (Attention Causal Convolutional Neural Network) model to increase the self-testing efficiency in subway switching stations was established in this paper. AC-CNN model is a deep learning prediction network based on attention mechanism, which can capture the causal convolutional layer of the time before and after relationship as the basic network structure, optimize the convergence process through residual connections, and fuse global and local attention representations through matrix multiplication in the attention mechanism module to reduce useful information loss. The algorithm used the one-hot encoding and sliding window feature engineering operations to preprocess the historical self-testing operation data and completed the prediction of the self-testing information of metro one-button switch station equipment through the deep analysis of AC-CNN. The on-site application results indicated that AC-CNN had greater advantages in terms of accuracy and operation time compared to those of existing BP neural networks, GBDT, and LSTM algorithms, which can significantly improve the work efficiency of subway staff on one-button switch station.

Detection of Rotating Stall Inception of Axial Compressors Based on Deep Dilated Causal Convolutional Neural Networks

The anomaly detection of outliers and fault diagnosis in the imperfect time-series data is crucial in aero-engine industry. Early detection of the rotating stall is significant for the active stabilization control of the axial compressors, because the rotating stall and surge can cause enormous vibratory stresses in the blades and gives rise to surge to limit the performance of compressors. An aerodynamic instability inception and short-length-scale periodic anomaly prior to stall onset known as stall inception in axial compressors is observed in aero-engine. Based on deep learning theory, this paper conducts the accurate and rapid detection of stall precursors based on deep neural network. The deep dilated causal convolutional neural network combined with logistic regression(LR) named as LR-WaveNet is applied to the detection and prediction of stall inception in the imperfect time-series data of axial compressors with the rotating stall. The LR-WaveNet model can implement fast anomaly detection and stall prediction in long-time term series data. Furthermore, a single LR-WaveNet can be trained to capture and learn the time-domain statistical characteristics of many different stall inception training data with equal fidelity. The trained LR-WaveNet model can rapidly detect the occurrence of anomaly point and predict the probability of rotating stall and surge in axial compressors as an early warning signal. By comparing with the time domain analysis and the wavelet analysis, the calculation results are represented with experimental data to show the effectiveness and feasibility of the stall detection approach based on LR-WaveNet. Especially, the LR-WaveNet can detect the irregular and imperfect stall inception and capture the small fault features when the characteristics of stall inception sigals or data are not very unobvious, irregular and imperfect. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —Note to Practitioners–This paper was motivated by the problem of detecting the stall and surge of axial compressor for aeroengine, but it also applies to other small fault diagnosis with imperfect data or time series prediction problem. The stall Inception is very difficult to detect for the reason of imperfect data with small fault or features. These existing detection methods of stall inception can only be applied under certain conditions, because they can only extract the pre-stall features partly. Meanwhile, the excessive number of compressor sensors is demanded in this method, so the application of these traditional methods are limited in practical situations in the stall inception detection. This paper suggests a new approach using Deep Dilated Causal Convolutional Neural Networks. The deep dilated causal convolutional neural network combined with logistic regression(LR) named as LR-WaveNet is applied to the detection and prediction of stall inception in the imperfect time-series data of axial compressors with the rotating stall. In this paper, the new approach is proved to be suitable and feasible to all the types of stall inception including the irregular stall inception with imperfect data. We then show how the deep learning method can be efficiently applied in the imperfect data situation with small fault and time series prediction. The experiments suggest that this approach is feasible. In future research, we will create the novel approach and theory for the detection of stall inception and small fault diagnostics with the imperfect data.

LightPRA: A Lightweight Temporal Convolutional Network for Automatic Physical Rehabilitation Exercise Assessment

Research evidence shows that physical rehabilitation exercises prescribed by medical experts can assist in restoring physical function, improving life quality, and promoting independence for physically disabled individuals. In response to the absence of immediate expert feedback on performed actions, developing a Human Action Evaluation (HAE) system emerges as a valuable automated solution, addressing the need for accurate assessment of exercises and guidance during physical rehabilitation. Previous HAE systems developed for the rehabilitation exercises have focused on developing models that utilize skeleton data as input to compute a quality score for each action performed by the patient. However, existing studies have focused on improving scoring performance while often overlooking computational efficiency. In this research, we propose LightPRA (Light Physical Rehabilitation Assessment) system, an innovative architectural solution based on a Temporal Convolutional Network (TCN), which harnesses the capabilities of dilated causal Convolutional Neural Networks (CNNs). This approach efficiently captures complex temporal features and characteristics of the skeleton data with lower computational complexity, making it suitable for real-time feedback provided on resource-constrained devices such as Internet of Things (IoT) devices and Edge computing frameworks. Through empirical analysis performed on the University of Idaho-Physical Rehabilitation Movement Data (UI-PRMD) and KInematic assessment of MOvement for remote monitoring of physical REhabilitation (KIMORE) datasets, our proposed LightPRA model demonstrates superior performance over several state-of-the-art approaches such as Spatial–Temporal Graph Convolutional Network (STGCN) and Long Short-Term Memory (LSTM)-based models in scoring human activity performance, while exhibiting lower computational cost and complexity.

Advanced Integration of Forecasting Models for Sustainable Load Prediction in Large-Scale Power Systems

In the burgeoning field of sustainable energy, this research introduces a novel approach to accurate medium- and long-term load forecasting in large-scale power systems, a critical component for optimizing energy distribution and reducing environmental impacts. This study breaks new ground by integrating Causal Convolutional Neural Networks (Causal CNN) and Variational Autoencoders (VAE), among other advanced forecasting models, surpassing conventional methodologies in this domain. Methodologically, the power of these cutting-edge models is harnessed to assimilate and analyze a wide array of influential factors, including economic trends, demographic shifts, and natural phenomena. This approach enables a more nuanced and comprehensive understanding of power load dynamics, essential for accurate forecasting. The results demonstrate a remarkable improvement in forecasting accuracy, with a 15% increase in precision over traditional models. Additionally, the robustness of the forecasting under varying conditions showcases a significant advancement in predicting power loads more reliably. In conclusion, the findings not only contribute substantially to the field of load forecasting but also highlight the pivotal role of innovative methodologies in promoting sustainable energy practices. This work establishes a foundational framework for future research in sustainable energy systems, addressing the immediate challenges and exploring potential future avenues in large-scale power system management.

A causal convolutional neural network for multi-subject motion modeling and generation

Inspired by the success of WaveNet in multi-subject speech synthesis, we propose a novel neural network based on causal convolutions for multi-subject motion modeling and generation. The network can capture the intrinsic characteristics of the motion of different subjects, such as the influence of skeleton scale variation on motion style. Moreover, after fine-tuning the network using a small motion dataset for a novel skeleton that is not included in the training dataset, it is able to synthesize high-quality motions with a personalized style for the novel skeleton. The experimental results demonstrate that our network can model the intrinsic characteristics of motions well and can be applied to various motion modeling and synthesis tasks.

Multi-view domain-adaptive representation learning for EEG-based emotion recognition

Current research suggests that there exist certain limitations in EEG emotion recognition, including redundant and meaningless time-frames and channels, as well as inter- and intra-individual differences in EEG signals from different subjects. To deal with these limitations, a Cross-attention-based Dilated Causal Convolutional Neural Network with Domain Discriminator (CADD-DCCNN) for multi-view EEG-based emotion recognition is proposed to minimize individual differences and automatically learn more discriminative emotion-related features. First, differential entropy (DE) features are obtained from the raw EEG signals using short-time Fourier transform (STFT). Second, each channel of the DE features is regarded as a view, and the attention mechanisms are utilized at different views to aggregate the discriminative affective information at the level of the time-frame of EEG. Then, a dilated causal convolutional neural network is employed to distill nonlinear relationships among different time frames. Next, a feature-level fusion is used to fuse features from multiple channels, aiming to explore the potential complementary information among different views and enhance the representational ability of the feature. Finally, to minimize individual differences, a domain discriminator is employed to generate domain-invariant features, which projects data from both the different domains into the same data representation space. We evaluated our proposed method on two public datasets, SEED and DEAP. The experimental results illustrate that our CADD-DCCNN method outperforms the SOTA methods.

A data-driven approach for railway in-train forces monitoring

Railway in-train forces are an essential element in assessing multiple aspects of rolling stocks. Conventional methods for obtaining the forces can be time-consuming and require significant investment in manpower and domain expertise, while only gathering the force data for specific service conditions one at a time. However, automatic train operation (ATO) systems can measure real-time information for trains and tracks by on-board and trackside devices, which could provide an opportunity for in-train forces monitoring. This paper presents a data-driven approach that uses ATO-measured data and a neural network model to monitor in-train forces under service conditions. To develop this approach, longitudinal train dynamics simulations (LTSs) for a freight train were conducted to establish the relationship between ATO measurements and in-train forces on specific couplers, which was embedded in a large amount of training data. After that, a specially developed self-attention-based causal convolutional neural network (SA-CNN) was employed to learn the underlying relationship and estimate the in-train force histories considering temporal dependencies. The comparative evaluation between the SA-CNN against four alternative neural network models revealed that the SA-CNN exhibits a slightly higher level of accuracy. Furthermore, the generalisation capability of the well-trained SA-CNN model was confirmed by numerical LTSs under four different service conditions. The results indicated that the data-driven approach has superior compatibility for arbitrarily combined inputs with significantly reduced computational time compared to LTSs. This approach holds the potential for achieving reliable in-situ monitoring of railway in-train forces, which is beneficial to both in-train force-related research and industrial applications.

Data-driven algorithm based on the scaled boundary finite element method and deep learning for the identification of multiple cracks in massive structures

Structural defect identification is a vital aspect of structural health monitoring used to assess the safety of engineering structures. However, quantitatively determining the dimensions of structural defects is often difficult. Therefore, this study presents an innovative data-driven algorithm that combines the scaled boundary finite element method (SBFEM) and a deep learning framework based on a dilated causal convolutional neural network (CNN) to identify crack-like defects in large-scale structures. The SBFEM is used to simulate different crack-like defects. Mesh generation is significantly simplified by a simple procedure that requires only changing the scale centre at the crack tip and the positions of the nodes at the crack opening. This minimises remeshing and enables simple generation of sufficient data to train the neural network. In addition, an absorbing boundary model based on Rayleigh damping is used to avoid computing the entire model when simulating wave propagation in massive structures. To ensure that sequential data remain ordered and to obtain a large receptive field without increasing the complexity of the neural network, a dilated causal CNN is employed in the deep learning framework. Therefore, more historical information is captured, and the complex mapping relationship between the echo signal and the crack information is efficiently learnt. The proposed model can accurately identify the number, location, and depth of cracks in massive structures. Moreover, it is robust to noise, which is demonstrated via numerical examples. Therefore, the proposed algorithm provides valuable insight into the detection and diagnosis of structural defects, which can ultimately improve the safety of engineering structures.

Electric Load Forecasting Based on Deep Ensemble Learning

Short-to-medium-term electric load forecasting is crucial for grid planning, transformation, and load scheduling for power supply departments. Various complex and ever-changing factors such as weather, seasons, regional economic structures, and enterprise production cycles exert uncontrollable effects on the electric grid load. While the causal convolutional neural network can significantly enhance long-term sequence prediction, it may suffer from problems such as vanishing gradients and overfitting due to extended time series. To address this issue, this paper introduces a new power load data anomaly detection method, which leverages a convolutional neural network (CNN) to extract temporal and spatial information from the load data. The features extracted are then processed using a bidirectional long short-term memory network (BiLSTM) to capture the temporal dependencies in the data more adeptly. An enhanced random forest (RF) classifier is employed for anomaly detection in electric load data. Furthermore, the paper proposes a new model framework for electricity load forecasting that combines a dilated causal convolutional neural network with ensemble learning. This combination addresses issues such as vanishing gradients encountered in causal convolutional neural networks with long time series. Extreme gradient boosting (XGBoost), category boosting (CATBoost), and light gradient boosting machine (LightGBM) models act as the base learners for ensemble modeling to comprehend deep cross-features, and the prediction results generated by ensemble learning serve as a new feature set for secondary ensemble modeling. The dilated convolutional neural network broadens the receptive field of the convolutional kernel. All acquired feature values are concatenated and input into the dilated causal convolutional neural network for training, achieving short-to-medium-term electric load forecasting. Experimental results indicate that compared to existing models, its root mean squared error (RMSE) and mean squared error (MSE) in short-term and mid-term electricity load forecasting are reduced by 4.96% and 12.31%, respectively, underscoring the efficacy of the proposed framework.

High-fidelity time-series data synthesis based on finite element simulation and data space mapping

Dynamic responses can provide rich information for supporting the entire life cycle of structures, and they can either be measured from actual structures or simulated using the finite element (FE) method. For the FE simulation, insufficient fidelity of simulation data can significantly affect the confidence of analysis results, while FE model updating methods can partially address this problem by reducing the simulation error. However, most FE model updating methods inevitably update the hyperparameters of FE models using sophisticated algorithms with high computational complexities. Thus, one question was raised: whether there is a mapping that can transfer the FE simulation data to the corresponding measurement data directly without performing FE model updating? To achieve this, we proposed a data synthesis method using FE simulation and deep learning space mapping, which can be used to synthesise high-fidelity dynamic responses excited by some unseen load patterns in the measurement. A Dilated Causal Convolutional Neural Network (DCCNN) was designed for realising space mapping. Vibration experiments were conducted on both an I-shaped steel beam and the corresponding FE model to establish datasets and test the proposed method. The quality of the synthetic data was analysed in both the time domain and the frequency domain. The accurate amplitudes, natural frequencies, and mode shapes of the synthetic data successfully demonstrate the effectiveness of the proposed high-fidelity data synthesis method.

Novel welch-transform based enhanced spectro-temporal analysis for cognitive microsleep detection using a single electrode EEG

The growing demand for semi-autonomous human–machine systems has led to an increased requirement for human fatigue detection. Direct and invasive approaches for microsleep detection include cognitive computing methods using Brain-Computer Interfaces (BCI). The contextual integration of multi-channel or heterogeneous signals for sleep staging remains a formidable challenge. In addition, the cost of acquiring many signals is significantly higher than that of acquiring a single signal. Consequently, researchers have recently attempted to utilize single-channel EEG over multi-channel acquisition systems for sleep staging. The Fast Fourier Transform (FFT) has been widely used in previous research findings for spectral analysis of complex time-series data streams. In contrast to the FFT, we utilize here, for the first time, the Welch Transform which can give higher stability in noise reduction for spectral analysis. Specifically, we provide a novel method to implement the short-time Welch transform (STWT) as an enhanced technique for the spectro-temporal analysis of single-electrode EEG signals. Further, our proposed model utilizes attention-based spatial and channel-wise inter-dependencies using a one-dimensional causal convolutional neural network (CNN) to extract contextual features automatically. Finally, we demonstrate an end-to-end proof of concept for our data extraction, adaptive data resampling, manual feature extraction, and deep-neural network-based modeling architecture. Comparative simulation results using the benchmark, maintenance of wake-fullness test (MWT) dataset for microsleep detection during automobile transportation, show that our proposed end-to-end system, utilizing novel STWT-based enhanced spectro-temporal analysis, outperforms current state-of-the-art methods, delivering 95% and 89% test accuracy for the case of temporal and spectral data inputs, respectively.

Dilated causal convolutional neural networks for forecasting zone airflow to estimate short-term energy consumption

Dilated causal convolutional neural networks for forecasting zone airflow to estimate short-term energy consumption

Wind Power Probability Density Prediction Based on Quantile Regression Model of Dilated Causal Convolutional Neural Network

Aiming at the wind power prediction problem, a wind power probability prediction method based on the quantile regression of a dilated causal convolutional neural network is proposed. With the developed model, the Adam stochastic gradient descent technique is utilized to solve the cavity parameters of the causal convolutional neural network under different quantile conditions and obtain the probability density distribution of wind power at various times within the following 200 hours. The presented method can obtain more useful information than conventional point and interval predictions. Moreover, a prediction of the future complete probability distribution of wind power can be realized. According to the actual data forecast of wind power in the PJM network in the United States, the proposed probability density prediction approach can not only obtain more accurate point prediction results, it also obtains the complete probability density curve prediction results for wind power. Compared with two other quantile regression methods, the developed technique can achieve a higher accuracy and smaller prediction interval range under the same confidence level.

Deep-Learning-Based Sequence Causal Long-Term Recurrent Convolutional Network for Data Fusion Using Video Data

The purpose of AI-Based schemes in intelligent systems is to advance and optimize system performance. Most intelligent systems adopt sequential data types derived from such systems. Realtime video data, for example, are continuously updated as a sequence to make necessary predictions for efficient system performance. The majority of deep-learning-based network architectures such as long short-term memory (LSTM), data fusion, two streams, and temporal convolutional network (TCN) for sequence data fusion are generally used to enhance robust system efficiency. In this paper, we propose a deep-learning-based neural network architecture for non-fix data that uses both a causal convolutional neural network (CNN) and a long-term recurrent convolutional network (LRCN). Causal CNNs and LRCNs use incorporated convolutional layers for feature extraction, so both architectures are capable of processing sequential data such as time series or video data that can be used in a variety of applications. Both architectures also have extracted features from the input sequence data to reduce the dimensionality of the data and capture the important information, and learn hierarchical representations for effective sequence processing tasks. We have also adopted a concept of series compact convolutional recurrent neural network (SCCRNN), which is a type of neural network architecture designed for processing sequential data combined by both convolutional and recurrent layers compactly, reducing the number of parameters and memory usage to maintain high accuracy. The architecture is challenge-able and suitable for continuously incoming sequence video data, and doing so allowed us to bring advantages to both LSTM-based networks and CNNbased networks. To verify this method, we evaluated it through a sequence learning model with network parameters and memory that are required in real environments based on the UCF-101 dataset, which is an action recognition data set of realistic action videos, collected from YouTube with 101 action categories. The results show that the proposed model in a sequence causal long-term recurrent convolutional network (SCLRCN) provides a performance improvement of at least 12% approximately or more to be compared with the existing models (LRCN and TCN).

Reverberation Suppression in Echocardiography Using a Causal Convolutional Neural Network

Reverberation Suppression in Echocardiography Using a Causal Convolutional Neural Network

High-precision DOA estimation for underwater acoustic signals based on sparsity adaptation

The direction of arrival (DOA) estimation technique is to obtain the direction information of the source when it reaches the array by processing and analyzing the received signal. In recent years, the DOA estimation of an array signal has been a research hotspot. For application scenarios with a small number of snapshots and a low signal-to-noise ratio, the compressive sensing theory has been commonly used to estimate the DOA of an array signal to achieve better estimation performance. However, the DOA estimation methods based on compressive sensing theory require information on source sparsity. Moreover, the influence of a complex underwater acoustic environment limits the accuracy of estimation algorithms. To address this limitation, this study proposes a high-precision DOA estimation model for underwater acoustic signals based on sparsity adaptation. The proposed model includes mainly two parts. In the first part, a source sparsity adaptive model based on a causal convolutional neural network is proposed. The model is used to address the constraint that the source sparsity should be known a priori when compressed sensing is used for DOA estimation. In the second part, a differential combination matching pursuit (DCMP) algorithm is adopted. First, a differentiated path filtering strategy is employed to reduce algorithm complexity and avoid the problem of invalid filtering. In addition, the combined optimization strategy is used to improve the prediction accuracy of the algorithm, providing an efficient error correction idea for the compressed sensing application to DOA estimation. The results of simulations conducted under seven different signal-to-noise ratios and using three different array types show that the proposed source sparsity adaptive model can reach an average prediction accuracy of 89.6%. In addition, compared with the other reconstruction algorithm accuracy, on the basis of ensuring low time complexity, the proposed DCMP algorithm can achieve an accuracy improvement of 9.99%–19.94% under seven different signal-to-noise ratio values. Moreover, the mean absolute error of the proposed DCMP algorithm is lower by approximately 0.05°–14° than those of the OMP and MMP algorithms.

Spatiotemporal causal convolutional network for forecasting hourly PM2.5 concentrations in Beijing, China

Air pollution in Northeastern Asia is a serious environmental problem, especially in China where PM2.5 levels are quite high. Accurate PM2.5 predictions are significant to environmental management and human health. Recently, deep learning has received increasing attention from relevant researchers. In this work, a spatiotemporal causal convolutional neural network (ST-CausalConvNet) for short-term PM2.5 prediction is proposed. The distinguishing characteristics of the proposed model is that the convolutions in the model architecture are causal, where an output at a certain time step is convolved only with elements from the same or earlier time steps in the previous layer. Accordingly, no information leakage is induced from the future to the past in this model. The spatial dependence between multiple monitoring stations was also considered in the model. Spatiotemporal correlation analysis was performed to select relevant information from monitoring stations that have a high relationship with the target station. The information from the target and related stations were then employed as the inputs and fed into the model. A case study from May 1, 2014 to April 30, 2015 in Beijing, China was conducted. The next hour PM2.5 concentration was predicted by the proposed model by using historical air quality and meteorological data from 36 monitoring stations. Experimental results show that the trends of the predicted PM2.5 concentrations and the observed values were consistent. The proposed method achieved a better prediction performance than the other three comparative models, namely artificial neural network (ANN), gated recurrent unit (GRU), and long short-term memory (LSTM). Furthermore, the effects of the important parameters and the model transferability were also conducted. We conclude that the proposed ST-CausalConvNet is a potential effective model for air pollution forecasting.

Real-time forecasting of petrol retail using dilated causal CNNs

The recent popularity of smart cities and smart homes has made the adoption of Internet of Things (IoT) devices ubiquitous. Most of these IoT devices are low-end devices with limited capabilities. For neural network based predictive models, the low processing power of connected things is a limitation when training them. In addition, it is still a common practice to deploy these models on cloud servers that possess dedicated high performance computing hardware. However, for IoT applications, it is not feasible to send voluminous raw data to the cloud or a remote backend server on account of high latency, information security concerns or lack of network coverage. In this work, we develop an integrated prediction system for a retail petrol station within the operational constraints of the IoT ecosystem. Our main contribution is the combination of the recent concepts of dilated convolution and the so-called causal convolution into the 1D dilated causal convolutional neural network for time-series prediction. This results in a significantly lightweight model with sound mathematical grounding. Empirical evaluation with a highly-relaxed grid search in the hyperparameter space shows an order-of-magnitude improvement over competing models, both in terms of predictive performance as well as computational cost while training. Compared to state-of-the-art models, our proposed model is able reduce the root mean squared error from 49.38 to 18.37, and training time from 93.07 to 10.21 s on the petrol sales prediction problem.