Length Of Historical Data Research Articles

PreSTNet: Pre-trained Spatio-Temporal Network for traffic forecasting

Traffic forecasting stands as a cornerstone in urban planning, yet existing methods mainly fall short in capturing long-term spatio-temporal patterns. To be specific, various exquisite modules are designed in current works, while bringing heavy computing costs with the length of historical data. Furthermore, any replacement or reintegration of modules could weaken final performance, and the whole model needs to be trained from scratch on different datasets. To address the above shortcomings, inspired by pre-training methodologies, we introduce a novel Pre-trained Spatio-Temporal Network (PreSTNet), a two-phase framework comprising pre-training and fine-tuning. In the first pre-training phase, PreSTNet undertakes a data masking and recovery task, which is facilitated through a dedicated long sequence embedding module and an encoder–decoder structure with temporal attention layers and graph convolution operators. The recovery task involves masking several data, forcing the model to capture correlations between unmasked data and finally recover the masked part, which facilitates the extraction of long-term spatio-temporal features from extensive data. The second phase involves freezing the learned parameters and replacing the decoder with a forecasting header. This header, designed as a meta-learning fusion module and spatio-temporal convolution layers, can integrate long-term and short-term traffic data, and is trained in the manner of supervised learning to suit the target task. Rigorous experiments on real-world datasets underscore the superiority of PreSTNet. Further case studies reveal that the pre-trained encoder, coupled with an intuitive linear regression, achieves comparable performance to advanced methods. These findings affirm the robust capabilities of the proposed PreSTNet in addressing the complexities of traffic forecasting.

Return levels of extreme European windstorms, their dependency on the North Atlantic Oscillation, and potential future risks

Abstract. Windstorms are the most damaging natural hazard across western Europe. Risk modellers are limited by the observational data record to only ∼ 60 years of comprehensive reanalysis data that are dominated by considerable inter-annual variability. This makes estimating return periods of rare events difficult and sensitive to the choice of the historical period used. This study proposes a novel statistical method for estimating wind gusts across Europe based on observed windstorm footprints. A good description of extreme wind speeds is obtained by assuming that gust speed peaks over threshold are distributed exponentially, i.e. a generalised Pareto distribution having a zero shape parameter. The threshold and tail scale parameter are estimated at each location and used to calculate estimates of the 10- and 200-year return levels. The North Atlantic Oscillation (NAO) is particularly important for modulating lower return levels and modulating the threshold, with a less detectable influence on rarer extremes and the tail scale parameter. The length of historical data required to have the lowest error in estimating return levels is quantified using both observed and simulated time series of the historical NAO. For reducing errors in estimating 200-year return levels of an independent 10-year period, a data catalogue of at least 20 years is required. For lower return levels the NAO has a stronger influence on estimated return levels, and so there is more variability in estimates. Using theoretical estimates of future NAO states, return levels are largely outside the historical uncertainty, indicating significant increases in risk potential from windstorms in the next 100 years. Our method presents a framework for assessing high-return-period events across a range of hazards without the additional complexities of a full catastrophe model.

A performance-based hybrid deep learning model for predicting TBM advance rate using Attention-ResNet-LSTM

The technology of tunnel boring machine (TBM) has been widely applied for underground construction worldwide; however, how to ensure the TBM tunneling process safe and efficient remains a major concern. Advance rate is a key parameter of TBM operation and reflects the TBM-ground interaction, for which a reliable prediction helps optimize the TBM performance. Here, we develop a hybrid neural network model, called Attention-ResNet-LSTM, for accurate prediction of the TBM advance rate. A database including geological properties and TBM operational parameters from the Yangtze River Natural Gas Pipeline Project is used to train and test this deep learning model. The evolutionary polynomial regression method is adopted to aid the selection of input parameters. The results of numerical experiments show that our Attention-ResNet-LSTM model outperforms other commonly-used intelligent models with a lower root mean square error and a lower mean absolute percentage error. Further, parametric analyses are conducted to explore the effects of the sequence length of historical data and the model architecture on the prediction accuracy. A correlation analysis between the input and output parameters is also implemented to provide guidance for adjusting relevant TBM operational parameters. The performance of our hybrid intelligent model is demonstrated in a case study of TBM tunneling through a complex ground with variable strata. Finally, data collected from the Baimang River Tunnel Project in Shenzhen of China are used to further test the generalization of our model. The results indicate that, compared to the conventional ResNet-LSTM model, our model has a better predictive capability for scenarios with unknown datasets due to its self-adaptive characteristic.

TBM performance prediction using LSTM-based hybrid neural network model: Case study of Baimang River tunnel project in Shenzhen, China

Accurately predicting tunnel boring machine (TBM) performance is beneficial for excavation efficiency enhancement and risk mitigation of TBM tunneling. In this paper, we develop a long short-term memory (LSTM) based hybrid intelligent model to predict two key TBM performance parameters (advance rate and cutterhead torque). The model combines the LSTM, BN, Dropout and Dense layers to process the raw data and improve the fitting quality. The features, including the ground formation properties, tunnel route curvature, tunnel location and TBM operational parameters, are divided into historical/real-time time-varying parameters, time-invariant parameters and historical/real-time output prediction data. The effectiveness of the proposed model is verified based on a large monitoring database of the Baimang River Tunnel Project in Shenzhen, south China. We then discuss the influence of the prediction mode, neural network structure and time division interval length of historical data on the prediction accuracy. The significance evaluation of input features shows that the historical output prediction has the largest influence on the prediction accuracy, and the influence of ground properties is secondary. It is also found that the correlations between input features and the output prediction are coincident with their interrelationships with the ground properties and ease of TBM excavation. Finally, it is found that the prediction results are most affected by the total propulsion force followed by the rotation speed of the cutterhead. The established model can provide useful guidance for construction personnel to roughly grasp the possible TBM status from the prediction results when adjusting the operational parameters.

Prediction of residential district heating load based on machine learning: A case study

Heating load prediction plays an important role in supporting the operation of a residential district energy station. To find out the most suitable prediction algorithm, seven popular machine learning algorithms are applied in a district heating load dataset from Shanghai, China. The data derived from electric power sensor, thermal sensor and meteorological sensor, weather forecast are combined into four input selections to evaluate model performance. Research on the length setting of prediction time horizon for prediction target and the length of historical time series data for input selection are also conducted to investigate corresponding performance changes. The results show that Support Vector Regression (SVR) performs best in Mean Average Percentage Error (MAPE) with 5.21%. XGBoost performs best in Root Mean Square Error (RMSE) with 59.75 kW. However, SVR, XGBoost and Long Short Term Memory neural network (LSTM) have similar performance. Further investigation shows that the consistent increase in the length of historical data sets does not help to improve the performance. The length of historical data sets is recommended to be 28 h. Meanwhile, the span of prediction horizon has slight influence on the performance. It is also found that model performance is primarily influenced by the weather forecasting data.

Dynamics of crime activities in the network of city community areas

Understanding criminal activities, their structure and dynamics are fundamental for designing tools for crime prediction that can also guide crime prevention. Here, we study crimes committed in city community areas based on police crime reports and demographic data for the City of Chicago collected over 16 consecutive years. Our goal is to understand how the network of city community areas shapes dynamics of criminal offenses and demographic characteristics of their inhabitants. Our results reveal the presence of criminal hot-spots and expose the dynamic nature of criminal activities. We identify the most influential features for forecasting the per capita crime rate in each community. Our results indicate that city community crime is driven by spatio-temporal dynamics since the number of crimes committed in the past among the spatial neighbors of each community area and in the community itself are the most important features in our predictive models. Moreover, certain urban characteristics appear to act as triggers for the spatial spreading of criminal activities. Using the k-Means clustering algorithm, we obtained three clearly separated clusters of community areas, each with different levels of crimes and unique demographic characteristics of the district’s inhabitants. Further, we demonstrate that crime predictive models incorporating both demographic characteristics of a community and its crime rate perform better than models relying only on one type of features. We develop predictive algorithms to forecast the number of future crimes in city community areas over the periods of one-month and one-year using varying sets of features. For one-month predictions using just the number of prior incidents as a feature, the critical length of historical data, τc, of 12 months arises. Using more than τc months ensures high accuracy of prediction, while using fewer months negatively impacts prediction quality. Using features based on demographic characteristics of the district’s inhabitants weakens this impact somewhat. We also forecast the number of crimes in each community area in the given year. Then, we study in which community area and over what period an increase in crime reduction funding in this area will yield the largest reduction of the crime in the entire city. Finally, we study and compare the performance of various supervised machine learning algorithms classifying reported crime incidents into the correct crime category. Using the temporal patterns of various crime categories improves the classification accuracy. The methodologies introduced here are general and can be applied to other cities for which data about criminal activities and demographics are available.

Mortality Forecasting: How Far Back Should We Look in Time?

Extrapolative methods are one of the most commonly-adopted forecasting approaches in the literature on projecting future mortality rates. It can be argued that there are two types of mortality models using this approach. The first extracts patterns in age, time and cohort dimensions either in a deterministic fashion or a stochastic fashion. The second uses non-parametric smoothing techniques to model mortality and thus has no explicit constraints placed on the model. We argue that from a forecasting point of view, the main difference between the two types of models is whether they treat recent and historical information equally in the projection process. In this paper, we compare the forecasting performance of the two types of models using Great Britain male mortality data from 1950–2016. We also conduct a robustness test to see how sensitive the forecasts are to the changes in the length of historical data used to calibrate the models. The main conclusion from the study is that more recent information should be given more weight in the forecasting process as it has greater predictive power over historical information.

Does Decision-Maker’s Intelligence Matter? Evidences From Distribution-Free Newsvendor Problem

Decision-makers’ intelligence is usually pre-assumed to be sufficient and contingent to satisfy various requirements in different decision support models. However, it can be hardly assured that decision-makers can be so intelligent as required in practical decision-making circumstances. This paper aims to demonstrate the influences of decision-makers’ intelligence level on decision-making models. Simulation and laboratory experiments are designed and carried out. In the simulation experiment, memory capacity and computation ability, two most important ingredients of intelligence, are introduced into the distribution-free newsvendor problem. Memory capacity is defined as the temporal length of historical data that newsvendor can memorize, and computation ability as intuitive complexity of the model applied to determine his optimal order. Then by setting up various intelligence levels, namely, different levels of memory capacity and computation ability, three typical ordering policies of various levels of intuitive complexity are simulated in the framework of the distribution-free newsvendor problem. Then, a laboratory experiment is executed. The ratio between the number of groups of historical demand and data analysis time length and the time length of order-making are defined as the comprehensive index of subject’s memory capacity and computation ability. The results of both the experiments demonstrate that the decision-maker’s memory capacity and computation ability have significant and positive influences on decision-maker’s expected profit. The results might present another piece of the reason that why decision-makers do not always act optimally.

Fast and Adaptive Cointegration Based Model for Forecasting High Frequency Financial Time Series

Cointegration is a long-run property of some non-stationary time series where a linear combination of those time series is stationary. This behaviour has been studied in finance because cointegration restrictions often improve forecasting. The vector error correction model (VECM) is a well-known econometric technique that characterises short-run variations of a set of cointegrated time series incorporating long-run relationships as an error correction term. VECM has been broadly used with low frequency time series. We aimed to adapt VECM to be used in finance with high frequency stream data. Cointegration relations change in time and therefore VECM parameters must be updated when new data is available. We studied how forecasting performance is affected when VECM parameters and the length of historical data used change in time. We observed that the number of cointegration relationships varies with the length of historical data used. Moreover, parameters that increased these relationships in time led to better forecasting performance. Our proposal, called an Adaptive VECM (AVECM) is to make a parameters grid search that maximises the number of cointegration relationships in the near past. To ensure the search can be executed fast enough, we used a distributed environment. The methodology was tested using four 10-s frequency time series of the Foreign Exchange market. We compared our proposal with ARIMA and the naive forecast of the random walk model. Numerical experiments showed that on average AVECM performed better than ARIMA and random walk. Additionally, AVECM significantly improved execution times with respect to its serial version.

Mortality Forecast: Local or Global?

Accurate future mortality forecasts are of fundamental importance as they ensure adequate pricing of mortality-linked insurance and financial products. Extrapolative methods are the most commonly adopted forecasting approach in the literature on projecting future mortality rates (see for example Clayton and Schifflers, 1987; Lee and Carter, 1992). There are generally two types of mortality models using the extrapolative approach. The first extracts patterns in age, time and cohort dimensions either in a deterministic fashion or a stochastic fashion (see for example Lee and Carter, 1992; Plat, 2009). The second uses non-parametric smoothing techniques to model mortality and thus has no explicit constraints placed on the model (see for example Currie, et al, 2004; Hyndman and Ullah, 2007). We argue that the main difference between the two types of models in terms of forecasting is the fact that, the former uses global information and the latter mainly uses local information. In this paper we conduct an investigation on the comparison of the forecasting performance of the two types of models using Great Britain male mortality data from 1950 to 2009. The paper assesses the accuracy of forecasts not only based on statistical measures but also take the randomness of residuals into account. A main conclusion from the study is that, local information seems to have greater predictive power over historical information so it should be given more weights in the forecasting process. We also conduct a robustness test to see how sensitive the forecasts are to the changes in the length of historical data used to calibrate the models.

Interbasin Water Transfer: Economic Water Quality-Based Model

The interbasin water transfer project is an alternative to balance the nonuniform temporal and spatial distribution of water resources and water demands, especially in arid and semi arid regions. A water transfer project can be executed if it is environmentally and economically justified. In this study, the feasibility of two interbasin water transfer projects from Karoon River in the western part of Iran to the central part of the country is investigated. An optimization model with an economic objective function to maximize the net benefit of the interbasin water transfer projects is developed. The planning horizon of the model is 23 years (the length of historical data); and it is solved using genetic algorithm. In order to consider environmental impacts of water transfer projects, a water quality simulation model has been used. Then, an Artificial Neural Network model is trained based on the simulation results of a river water quality model in order to be coupled with the optimization model. The outputs of the optimization model are the value of economic gain of the sending (Karoon) basin to offset the loss of agricultural income and environmental costs. The optimal polices for water transfer during the planning horizon has been generated using the coupled simulation-optimization model. Then, operating rules are developed using a K Nearest Neighborhood model for the real time water transfer operation. The results show the significant value of using the proposed algorithm and economic evaluation for water transfer projects.

Analogue correction of errors based on seasonal climatic predictable components and numerical experiments

Based on the idea of using historical-analogue information to predict the prediction errors of model, a new method named analogue correction of errors by predictable component （FACEPC） was developed. This method is adopted to identify the predicable components for which the prediction result is relatively not quite sensitive to the initial values. And then for predicable components, an associated scheme is chosen for historical-analogue selection and error correction. This method was further applied to experiments on operational seasonal prediction model of National Climate Center. By selecting suitable analogues and prediction schemes for different regions, the results from cross-validation indicate that the predictive skill scores of summer precipitation and circulation have got significant improvement relative to systematic error correction, which looks more obvious in ENSO episodes and over regions with more predictable components. Especially, the skill scores over China area have also been clearly improved, exhibiting its potential application perspective to operational seasonal prediction. Besides, preliminary sensitive experiments show that the FACEPC-based predictions are also obviously influenced by the analogue-selected factors and the length of historical data.

K odhadu volatility finančních řad při oceňování derivátů

Estimation of volatility of financial time series plays a crucial role in pricing derivatives. Volatility is often estimated from historical data; however, it is well known that volatility varies in time. We propose a method to choose a suitable length of historical data to estimate contemporary volatility. The method is based on adaptation of a procedure used in statistical quality control - a hypothesis, that data contains a changepoint of volatility, is tested and if the test gives a positive answer, the changepoint is estimated. Then, a period of data where no changepoint is statistically significant is used to estimate contemporary volatility. The approach is illustrated on an analysis of CZK/EUR exchange rates.

GENERATION OF DAILY AND HOURLY SOLAR RADIATION VALUES FOR MODELINGWATER QUALITY IN AQUACULTURE PONDS

A stochastic model has been developed for generation of daily and hourly solar radiation values that can be used as inputs in a water quality model for aquaculture ponds. The daily solar radiation values are generated based on the monthly probability distribution of the daily clearness index. The monthly probability distributions are obtained from an incomplete historical daily solar radiation data set collected from fish pond sites. The hourly solar radiation values are estimated by breaking down the generated daily value using an empirical equation from the literature. The model has been applied to locations in Thailand, Honduras, and Rwanda. The length of historical data for the different locations ranged between six and eight years. The generated values show good agreement with the measured data. This model can be used to generate solar radiation values for locations having scant historical information.

Free-radical polymerizations associated with the Trommsdorff effect under semibatch reactor conditions. IV. On-line inferential-state estimation

A moving-horizon inferential-state estimation technique is described which uses simulated “experimental” data on temperature and viscosity to study bulk polymerization of free-radical systems. The short-term predictive capability of this technique is found to be quite good. A considerable amount of ringing (oscillations between the lower and upper bounds) is observed in the values of the estimated parameters which can be reduced significantly by narrowing down the range of parameter values or by including longer horizons in parameter estimation. Short-range prediction of viscosity was also found to be good. The model-calculated values of monomer conversion and molecular weights were found to be quite satisfactory in the entire range of operation. The long-term predictions of the model using the estimated parameters may or may not be accurate depending on the length of historical data used in the prediction. However, periodic use of state-variable estimation based on all the data up to that time, followed by the determination of the optimal temperature history in the future, could be a feasible strategy for experimental on-line optimizing control of bulk free-radical polymerizations which exhibit significant amounts of the Trommsdorff effect. © 1997 John Wiley & Sons, Inc. J Appl Polym Sci 64: 1861–1877, 1997