Days Of Training Data Research Articles

Tutorial on time series prediction using 1D-CNN and BiLSTM: A case example of peak electricity demand and system marginal price prediction

Although research on time series prediction based on deep learning is being actively carried out in various industries, deep learning technology still has a high entry barrier for researchers who have not majored in computer science. This paper presents a tutorial on time series prediction using a deep learning-based model. The entire process of time series data prediction is presented—from data collection to evaluation of prediction results. The details of each step are shown through a case example of predicting peak electricity demand and system marginal price of Jeju Island in Korea using the 1D-CNN and BiLSTM model. In Jeju Island, the proportion of renewable energy in the total power generation capacity is increased to 67% in 2021, requiring more accurate electricity demand forecasts. Therefore, using 808 days of training data from February 2018, electricity demand and SMP for the next 21 days were predicted. To make it easier for readers to follow, the example uses only open public data, and the entire Python source code is shared via a GitHub repository. The prediction error calculated by WRMSSE showed 0.42 in electricity demand and 0.63 in SMP max. A WRMSSE value less than one means that the forecast was relatively good, that is, better than naïve forecasting. This tutorial is not limited to the energy industry but can be utilized for any application requiring time-series data prediction. This article is expected to be of great help to researchers who need to understand the process of time series prediction using deep learning and use it for application in their industry.

Reinforcement Learning for Intrusion Detection: More Model Longness and Fewer Updates

Several works have used machine learning techniques for network-based intrusion detection over the past few years. While proposed schemes have been able to provide high detection accuracies, they do not adequately handle the changes in network traffic behavior as time passes. Researchers often assume that model updates can be performed periodically as needed, although this is not easily feasible in real-world scenarios. This paper proposes a new intrusion detection model based on a reinforcement learning approach that aims to support extended periods without model updates. The proposal is divided into two strategies. First, it applies machine learning scheme as a reinforcement learning task to long-term learning -maintaining high reliability and high classification accuracies over time. Second, model updates are performed using a transfer learning technique coped with a sliding window mechanism that significantly decreases the need for computational resources and human intervention. Experiments performed using a new dataset spanning 8TB of data and four years of real network traffic indicate that current approaches in the literature cannot handle the evolving behavior of network traffic. Nevertheless, the proposed technique without periodic model updates achieves similar accuracy rates to traditional detection schemes implemented with semestral updates. In the case of performing periodic updates on our proposed model, it decreases the false positives up to 8%, false negatives up to 34%, with an accuracy variation up to only 6%, while demanding only seven days of training data and almost five times fewer computational resources when compared to traditional approaches.

Model selection, adaptation, and combination for transfer learning in wind and photovoltaic power forecasts

There is recent interest in using model hubs – a collection of pre-trained models – in computer vision tasks. To employ a model hub, we first select a source model and then adapt the model for the target to compensate for differences. There still needs to be more research on model selection and adaption for renewable power forecasts. In particular, none of the related work examines different model selection and adaptation strategies for neural network architectures. Also, none of the current studies investigates the influence of available training samples and considers seasonality in the evaluation. We close these gaps by conducting the first thorough experiment for model selection and adaptation for transfer learning in renewable power forecast, adopting recent developments from the field of computer vision on 667 wind and photovoltaic parks from six datasets. We simulate different amounts of training samples for each season to calculate informative forecast errors. We examine the marginal likelihood and forecast error for model selection for those amounts. Furthermore, we study four adaption strategies. As an extension of the current state of the art, we utilize a Bayesian linear regression for forecasting the response based on features extracted from a neural network. This approach outperforms the baseline with only seven days of training data and shows that fine-tuning is not beneficial with less than three months of data. We further show how combining multiple models through ensembles can significantly improve the model selection and adaptation approach such that we have a similar mean error with only 30 days of training data which is otherwise only possible with an entire year of training data. We achieve a mean error of 9.8 and 14 percent for the most realistic dataset for PV and wind with only seven days of training data.

Probabilistic Intraday PV Power Forecast Using Ensembles of Deep Gaussian Mixture Density Networks

There is a growing interest of estimating the inherent uncertainty of photovoltaic (PV) power forecasts with probability forecasting methods to mitigate accompanying risks for system operators. This study aims to advance the field of probabilistic PV power forecast by introducing and extending deep Gaussian mixture density networks (MDNs). Using the sum of the weighted negative log likelihood of multiple Gaussian distributions as a minimizing objective, MDNs can estimate flexible uncertainty distributions with nearly all neural network structures. Thus, the advantages of advances in machine learning, in this case deep neural networks, can be exploited. To account for the epistemic (e.g., model) uncertainty as well, this study applies two ensemble approaches to MDNs. This is particularly relevant for industrial applications, as there is often no extensive (manual) adjustment of the forecast model structure for each site, and only a limited amount of training data are available during commissioning. The results of this study suggest that already seven days of training data are sufficient to generate significant improvements of 23.9% in forecasting quality measured by normalized continuous ranked probability score (NCRPS) compared to the reference case. Furthermore, the use of multiple Gaussian distributions and ensembles increases the forecast quality relatively by up to 20.5% and 19.5%, respectively.

Can Fake News Detection Models Maintain the Performance through Time? A Longitudinal Evaluation of Twitter Publications

The negative impact of false information on social networks is rapidly growing. Current research on the topic focused on the detection of fake news in a particular context or event (such as elections) or using data from a short period of time. Therefore, an evaluation of the current proposals in a long-term scenario where the topics discussed may change is lacking. In this work, we deviate from current approaches to the problem and instead focus on a longitudinal evaluation using social network publications spanning an 18-month period. We evaluate different combinations of features and supervised models in a long-term scenario where the training and testing data are ordered chronologically, and thus the robustness and stability of the models can be evaluated through time. We experimented with 3 different scenarios where the models are trained with 15-, 30-, and 60-day data periods. The results show that detection models trained with word-embedding features are the ones that perform better and are less likely to be affected by the change of topics (for example, the rise of COVID-19 conspiracy theories). Furthermore, the additional days of training data also increase the performance of the best feature/model combinations, although not very significantly (around 2%). The results presented in this paper build the foundations towards a more pragmatic approach to the evaluation of fake news detection models in social networks.

Application of LSTM Networks for Water Demand Prediction in Optimal Pump Control

Every morning, water suppliers need to define their pump schedules for the next 24 h for drinking water production. Plans must be designed in such a way that drinking water is always available and the amount of unused drinking water pumped into the network is reduced. Therefore, operators must accurately estimate the next day’s water consumption profile. In real-life applications with standard consumption profiles, some expert system or vector autoregressive models are used. Still, in recent years, significant improvements for time series prediction have been achieved through special deep learning algorithms called long short-term memory (LSTM) networks. This paper investigates the applicability of LSTM models for water demand prediction and optimal pump control and compares LSTMs against other methods currently used by water suppliers. It is shown that LSTMs outperform other methods since they can easily integrate additional information like the day of the week or national holidays. Furthermore, the online- and transfer-learning capabilities of the LSTMs are investigated. It is shown that LSTMs only need a couple of days of training data to achieve reasonable results. As the focus of the paper is on the real-world application of LSTMs, data from two different water distribution plants are used for benchmarking. Finally, it is shown that the LSTMs significantly outperform the system currently in operation.

Improving wind vector predictions for modelling of atmospheric dispersion during Seveso-type accidents

In case of a major accident involving airborne emissions of harmful gases, a temporary portable meteorological station may be used to improve atmospheric dispersion modelling (ADM) for protection of people and the environment. While the meteorological station provides signal values in real time, ADM results for the future are of particular interest for planning purposes. It is possible to use the current measured value as a future input to the ADM but it is suboptimal. It is also possible to use model output statistics (MOS) to predict the future local weather information from numerical weather prediction (NWP) models, while available operational NWP models are in general too coarse to be used directly in fine-resolution ADM. MOS models are obtained through machine learning and the training data sets in most traditional uses of MOS are big, which is beneficial for modelling. We envision using MOS in an emergency and for a location of a temporary meteorological station. We use windowing for online data selection to explore its accuracy when the amount of available training data is very limited, which is expected in an emergency situation. We show that MOS for wind vector with 1 day of training data greatly improves on the numerical weather predictions and the persistence model, so its use in such an emergency would be advantageous.

Inhibition of Long-Term Variability in Decoding Forelimb Trajectory Using Evolutionary Neural Networks With Error-Correction Learning.

Objective: In brain machine interfaces (BMIs), the functional mapping between neural activities and kinematic parameters varied over time owing to changes in neural recording conditions. The variability in neural recording conditions might result in unstable long-term decoding performance. Relevant studies trained decoders with several days of training data to make them inherently robust to changes in neural recording conditions. However, these decoders might not be robust to changes in neural recording conditions when only a few days of training data are available. In time-series prediction and feedback control system, an error feedback was commonly adopted to reduce the effects of model uncertainty. This motivated us to introduce an error feedback to a neural decoder for dealing with the variability in neural recording conditions.Approach: We proposed an evolutionary constructive and pruning neural network with error feedback (ECPNN-EF) as a neural decoder. The ECPNN-EF with partially connected topology decoded the instantaneous firing rates of each sorted unit into forelimb movement of a rat. Furthermore, an error feedback was adopted as an additional input to provide kinematic information and thus compensate for changes in functional mapping. The proposed neural decoder was trained on data collected from a water reward-related lever-pressing task for a rat. The first 2 days of data were used to train the decoder, and the subsequent 10 days of data were used to test the decoder.Main Results: The ECPNN-EF under different settings was evaluated to better understand the impact of the error feedback and partially connected topology. The experimental results demonstrated that the ECPNN-EF achieved significantly higher daily decoding performance with smaller daily variability when using the error feedback and partially connected topology.Significance: These results suggested that the ECPNN-EF with partially connected topology could cope with both within- and across-day changes in neural recording conditions. The error feedback in the ECPNN-EF compensated for decreases in decoding performance when neural recording conditions changed. This mechanism made the ECPNN-EF robust against changes in functional mappings and thus improved the long-term decoding stability when only a few days of training data were available.

Neural Decoding Forelimb Trajectory Using Evolutionary Neural Networks with Feedback-Error-Learning Schemes.

Changes in the functional mapping between neural activities and kinematic parameters over time poses a challenge to current neural decoder of brain machine interfaces (BMIs). Traditional decoders robust to changes in functional mappings required many day's training data. The decoder may not be robust when it was trained by data from only few days. Therefore, a decoder should be trained to handle a variety of neural-to-kinematic mappings using limited training data. We proposed an evolutionary neural network with error feedback, ECPNN-EF, as a neural decoder, that considered the previous error as an input to the decoder in order to improve the robustness. The decoder was validated to reconstruct rat's forelimb movement in a water-reward lever-pressing task. Two days of data were only used to train the decoder while ten days of data were used to test the decoder. The results showed that the performance of ECPNN-EF was significantly higher than that of standard recurrent neural network without error feedback, which was commonly used in BMI. This suggested that ECPNN-EF trained with few days of training data can be robust to changes in functional mappings.

Early detection of spamming accounts in large-Scale service provider networks

We present ErDOS — an algorithm for the Early Detection Of Spamming accounts. The detection approach implemented by ErDOS combines content-based labelling and features based on inter-account communication patterns. We define new account features, based on the ratio between the numbers of sent and received emails, the distribution of emails received from different accounts, and the topological features of the network induced by inter-account communication. We also present ErDOS-LVS — a variant of ErDOS that targets the detection of low-volume spammers. The empirical evaluation of both detectors is based on a real-life data set collected by an email service provider, much larger than data sets previously used for outgoing spam detection research. It establishes that both are able to provide effective early detection of the spammers population, that is, they identify these accounts as spammers before they are detected as such by a content-based detector. Moreover, both detectors require only a single day of training data for providing a high-quality list of suspect accounts.

The U.S. National Blend of Models for Statistical Postprocessing of Probability of Precipitation and Deterministic Precipitation Amount

AbstractThe U.S. National Blend of Models provides statistically postprocessed, high-resolution multimodel ensemble guidance, providing National Weather Service forecasters with a calibrated, downscaled starting point for producing digital forecasts.Forecasts of 12-hourly probability of precipitation (POP12) over the contiguous United States are produced as follows: 1) Populate the forecast and analyze cumulative distribution functions (CDFs) to be used later in quantile mapping. Were every grid point processed without benefit of data from other points, 60 days of training data would likely be insufficient for estimating CDFs and adjusting the errors in the forecast. Accordingly, “supplemental” locations were identified for each grid point, and data from the supplemental locations were used to populate the forecast and analyzed CDFs used in the quantile mapping. 2) Load the real-time U.S. and Environment Canada (now known as Environment and Climate Change Canada) global deterministic and ensemble forecasts, interpolated to ⅛°. 3) Using CDFs from the past 60 days of data, apply a deterministic quantile mapping to the ensemble forecasts. 4) Dress the resulting ensemble with random noise. 5) Generate probabilities from the ensemble relative frequency. 6) Spatially smooth the forecast using a Savitzky–Golay smoother, applying more smoothing in flatter areas.Forecasts of 6-hourly quantitative precipitation (QPF06) are more simply produced as follows: 1) Form a grand ensemble mean, again interpolated to ⅛°. 2) Quantile map the mean forecast using CDFs of the ensemble mean and analyzed distributions. 3) Spatially smooth the field, similar to POP12.Results for spring 2016 are provided, demonstrating that the postprocessing improves POP12 reliability and skill, as well as the deterministic forecast bias, while maintaining sharpness and spatial detail.

Predicting Daily Activities From Egocentric Images Using Deep Learning.

We present a method to analyze images taken from a passive egocentric wearable camera along with the contextual information, such as time and day of week, to learn and predict everyday activities of an individual. We collected a dataset of 40,103 egocentric images over a 6 month period with 19 activity classes and demonstrate the benefit of state-of-the-art deep learning techniques for learning and predicting daily activities. Classification is conducted using a Convolutional Neural Network (CNN) with a classification method we introduce called a late fusion ensemble. This late fusion ensemble incorporates relevant contextual information and increases our classification accuracy. Our technique achieves an overall accuracy of 83.07% in predicting a person's activity across the 19 activity classes. We also demonstrate some promising results from two additional users by fine-tuning the classifier with one day of training data.

Verification and Calibration of Neighborhood and Object-Based Probabilistic Precipitation Forecasts from a Multimodel Convection-Allowing Ensemble

Abstract Neighborhood and object-based probabilistic precipitation forecasts from a convection-allowing ensemble are verified and calibrated. Calibration methods include logistic regression, one- and two-parameter reliability-based calibration, and cumulative distribution function (CDF)-based bias adjustment. Newly proposed object-based probabilistic forecasts for the occurrence of a forecast object are derived from the percentage of ensemble members with a matching object. Verification and calibration of single- and multimodel subensembles are performed to explore the effect of using multiple models. The uncalibrated neighborhood-based probabilistic forecasts have skill minima during the afternoon convective maximum. Calibration generally improves the skill, especially during the skill minima, resulting in positive skill. In general all calibration methods perform similarly, with a slight advantage of logistic regression (one-parameter reliability based) calibration for 1-h (6 h) accumulations. The uncalibrated object-based probabilistic forecasts are, in general, less skillful than the uncalibrated neighborhood-based probabilistic forecasts. Object-based calibration also results in positive skill at all lead times. For object-based calibration the skill is significantly different among the calibration methods, with the logistic regression performing the best and CDF-based bias adjustment performing the worst. For both the neighborhood and object-based probabilistic forecasts, the impact of using 10 or 25 days of training data for calibration is generally small and is most significant for the two-parameter reliability-based method. An uncalibrated Advanced Research Weather Research and Forecasting Model (ARW-WRF) subensemble is significantly more skillful than an uncalibrated WRF Nonhydrostatic Mesoscale Model (NMM) subensemble. The difference is reduced by calibration. The multimodel subensemble only shows an advantage for the neighborhood-based forecasts beyond 1-day lead time and shows no advantage for the object-based forecasts.

Probabilistic Forecast Calibration Using ECMWF and GFS Ensemble Reforecasts. Part I: Two-Meter Temperatures

Abstract Recently, the European Centre for Medium-Range Weather Forecasts (ECMWF) produced a reforecast dataset for a 2005 version of their ensemble forecast system. The dataset consisted of 15-member reforecasts conducted for the 20-yr period 1982–2001, with reforecasts computed once weekly from 1 September to 1 December. This dataset was less robust than the daily reforecast dataset produced for the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS), but it utilized a much higher-resolution, more recent model. This manuscript considers the calibration of 2-m temperature forecasts using these reforecast datasets as well as samples of the last 30 days of training data. Nonhomogeneous Gaussian regression was used to calibrate forecasts at stations distributed across much of North America. Significant observations included the following: (i) although the “raw” GFS forecasts (probabilities estimated from ensemble relative frequency) were commonly unskillful as measured in continuous ranked probability skill score (CRPSS), after calibration with a 20-yr set of weekly reforecasts their skill exceeded that of the raw ECMWF forecasts; (ii) statistical calibration using the 20-yr weekly ECMWF reforecast dataset produced a large improvement relative to the raw ECMWF forecasts, such that the ∼4–5-day calibrated reforecast-based product had a CRPSS as large as a 1-day raw forecast; (iii) a calibrated multimodel GFS/ECMWF forecast trained on 20-yr weekly reforecasts was slightly more skillful than either the individual calibrated GFS or ECMWF reforecast products; (iv) approximately 60%–80% of the improvement from calibration resulted from the simple correction of time-averaged bias; (v) improvements were generally larger at locations where the forecast skill was originally lower, and these locations were commonly found in regions of complex terrain; (vi) the past 30 days of forecasts were adequate as a training dataset for short-lead forecasts, but longer-lead forecasts benefited from more training data; and (vii) a small but consistent improvement was produced by calibrating GFS forecasts using the full 25-yr, daily reforecast training dataset versus the subsampled, 20-yr weekly training dataset.

Probabilistic Forecast Calibration Using ECMWF and GFS Ensemble Reforecasts. Part II: Precipitation

Abstract As a companion to Part I, which discussed the calibration of probabilistic 2-m temperature forecasts using large training datasets, Part II discusses the calibration of probabilistic forecasts of 12-hourly precipitation amounts. Again, large ensemble reforecast datasets from the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Global Forecast System (GFS) were used for testing and calibration. North American Regional Reanalysis (NARR) 12-hourly precipitation analysis data were used for verification and training. Logistic regression was used to perform the calibration, with power-transformed ensemble means and spreads as predictors. Forecasts were produced and validated for every NARR grid point in the conterminous United States (CONUS). Training sample sizes were increased by including data from 10 nearby grid points with similar analyzed climatologies. “Raw” probabilistic forecasts from each system were considered, in which probabilities were set according to ensemble relative frequency. Calibrated forecasts were also considered based on three amounts of training data: the last 30 days of forecasts (available for 2005 only), weekly reforecasts during 1982–2001, and daily reforecasts during 1979–2003 (GFS only). Several main results were found. (i) Raw probabilistic forecasts from the ensemble prediction systems’ relative frequency possessed little or negative skill when skill was computed with a version of the Brier skill score (BSS) that does not award skill solely on the basis of differences in climatological probabilities among samples. ECMWF raw forecasts had larger skills than GFS raw forecasts. (ii) After calibration with weekly reforecasts, ECMWF forecasts were much improved in reliability and were moderately skillful. Similarly, GFS-calibrated forecasts were much more reliable, albeit somewhat less skillful. Nonetheless, GFS-calibrated forecasts were much more skillful than ECMWF raw forecasts. (iii) The last 30 days of training data produced calibrated forecasts of light-precipitation events that were nearly as skillful as those with weekly reforecast data. However, for higher precipitation thresholds, calibrated forecasts using the weekly reforecast datasets were much more skillful, indicating the importance of large sample size for the calibration of unusual and rare events. (iv) Training with daily GFS reforecast data provided calibrated forecasts with a skill only slightly improved relative to that from the weekly data.