Search results for: LSTMs

Authors: Farzad Hosseini Hossein Abadi, Cristina Prieto Sierra, Cesar Álvarez Díaz

Abstract:

Predicting streamflow accurately in flashy catchments prone to floods is a major research and operational challenge in hydrological modeling. Recent advancements in deep learning, particularly Long Short-Term Memory (LSTM) networks, have shown to be promising in achieving accurate hydrological predictions at daily and hourly time scales. In this work, a multi-timescale LSTM (MTS-LSTM) network was applied to the context of regional hydrological predictions at an hourly time scale in flashy catchments. The case study includes 40 catchments allocated in the Basque Country, north of Spain. We explore the impact of hyperparameters on the performance of streamflow predictions given by regional deep learning models through systematic hyperparameter tuning - where optimal regional values for different catchments are identified. The results show that predictions are highly accurate, with Nash-Sutcliffe (NSE) and Kling-Gupta (KGE) metrics values as high as 0.98 and 0.97, respectively. A principal component analysis reveals that a hyperparameter related to the length of the input sequence contributes most significantly to the prediction performance. The findings suggest that input sequence lengths have a crucial impact on the model prediction performance. Moreover, employing catchment-scale analysis reveals distinct sequence lengths for individual basins, highlighting the necessity of customizing this hyperparameter based on each catchment’s characteristics. This aligns with well known “uniqueness of the place” paradigm. In prior research, tuning the length of the input sequence of LSTMs has received limited focus in the field of streamflow prediction. Initially it was set to 365 days to capture a full annual water cycle. Later, performing limited systematic hyper-tuning using grid search, revealed a modification to 270 days. However, despite the significance of this hyperparameter in hydrological predictions, usually studies have overlooked its tuning and fixed it to 365 days. This study, employing a simultaneous systematic hyperparameter tuning approach, emphasizes the critical role of input sequence length as an influential hyperparameter in configuring LSTMs for regional streamflow prediction. Proper tuning of this hyperparameter is essential for achieving accurate hourly predictions using deep learning models.

Keywords: LSTMs, streamflow, hyperparameters, hydrology

Procedia PDF Downloads 19

Authors: Vikrant Gupta, Amrit Goswami

Abstract:

The fixed income market forms the basis of the modern financial market. All other assets in financial markets derive their value from the bond market. Owing to its over-the-counter nature, corporate bonds have relatively less data publicly available and thus is researched upon far less compared to Equities. Bond price prediction is a complex financial time series forecasting problem and is considered very crucial in the domain of finance. The bond prices are highly volatile and full of noise which makes it very difficult for traditional statistical time-series models to capture the complexity in series patterns which leads to inefficient forecasts. To overcome the inefficiencies of statistical models, various machine learning techniques were initially used in the literature for more accurate forecasting of time-series. However, simple machine learning methods such as linear regression, support vectors, random forests fail to provide efficient results when tested on highly complex sequences such as stock prices and bond prices. hence to capture these intricate sequence patterns, various deep learning-based methodologies have been discussed in the literature. In this study, a recurrent neural network-based deep learning model using long short term networks for prediction of corporate bond prices has been discussed. Long Short Term networks (LSTM) have been widely used in the literature for various sequence learning tasks in various domains such as machine translation, speech recognition, etc. In recent years, various studies have discussed the effectiveness of LSTMs in forecasting complex time-series sequences and have shown promising results when compared to other methodologies. LSTMs are a special kind of recurrent neural networks which are capable of learning long term dependencies due to its memory function which traditional neural networks fail to capture. In this study, a simple LSTM, Stacked LSTM and a Masked LSTM based model has been discussed with respect to varying input sequences (three days, seven days and 14 days). In order to facilitate faster learning and to gradually decompose the complexity of bond price sequence, an Empirical Mode Decomposition (EMD) has been used, which has resulted in accuracy improvement of the standalone LSTM model. With a variety of Technical Indicators and EMD decomposed time series, Masked LSTM outperformed the other two counterparts in terms of prediction accuracy. To benchmark the proposed model, the results have been compared with traditional time series models (ARIMA), shallow neural networks and above discussed three different LSTM models. In summary, our results show that the use of LSTM models provide more accurate results and should be explored more within the asset management industry.

Keywords: bond prices, long short-term memory, time series forecasting, empirical mode decomposition

Procedia PDF Downloads 106

Authors: Konstantinos Perifanos, Eirini Florou, Dionysis Goutsos

Abstract:

This paper presents and benchmarks a number of end-to-end Deep Learning based models for metaphor detection in Greek. We combine Convolutional Neural Networks and Recurrent Neural Networks with representation learning to bear on the metaphor detection problem for the Greek language. The models presented achieve exceptional accuracy scores, significantly improving the previous state-of-the-art results, which had already achieved accuracy 0.82. Furthermore, no special preprocessing, feature engineering or linguistic knowledge is used in this work. The methods presented achieve accuracy of 0.92 and F-score 0.92 with Convolutional Neural Networks (CNNs) and bidirectional Long Short Term Memory networks (LSTMs). Comparable results of 0.91 accuracy and 0.91 F-score are also achieved with bidirectional Gated Recurrent Units (GRUs) and Convolutional Recurrent Neural Nets (CRNNs). The models are trained and evaluated only on the basis of training tuples, the related sentences and their labels. The outcome is a state-of-the-art collection of metaphor detection models, trained on limited labelled resources, which can be extended to other languages and similar tasks.

Keywords: metaphor detection, deep learning, representation learning, embeddings

Procedia PDF Downloads 117

Authors: Rohan Bhasin

Abstract:

Agrammatism in non-fluent Aphasia Cases can be defined as a language disorder wherein a patient can only use content words ( nouns, verbs and adjectives ) for communication and their speech is devoid of functional word types like conjunctions and articles, generating speech of with extremely rudimentary grammar . Past approaches involve Speech Therapy of some order with conversation analysis used to analyse pre-therapy speech patterns and qualitative changes in conversational behaviour after therapy. We describe this approach as a novel method to generate functional words (prepositions, articles, ) around content words ( nouns, verbs and adjectives ) using a combination of Natural Language Processing and Deep Learning algorithms. The applications of this approach can be used to assist communication. The approach the paper investigates is : LSTMs or Seq2Seq: A sequence2sequence approach (seq2seq) or LSTM would take in a sequence of inputs and output sequence. This approach needs a significant amount of training data, with each training data containing pairs such as (content words, complete sentence). We generate such data by starting with complete sentences from a text source, removing functional words to get just the content words. However, this approach would require a lot of training data to get a coherent input. The assumptions of this approach is that the content words received in the inputs of both text models are to be preserved, i.e, won't alter after the functional grammar is slotted in. This is a potential limit to cases of severe Agrammatism where such order might not be inherently correct. The applications of this approach can be used to assist communication mild Agrammatism in non-fluent Aphasia Cases. Thus by generating these function words around the content words, we can provide meaningful sentence options to the patient for articulate conversations. Thus our project translates the use case of generating sentences from content-specific words into an assistive technology for non-Fluent Aphasia Patients.

Keywords: aphasia, expressive aphasia, assistive algorithms, neurology, machine learning, natural language processing, language disorder, behaviour disorder, sequence to sequence, LSTM

Procedia PDF Downloads 140

Authors: Kostas Kouvaris, Anastasia Eleftheriou, Georgios A. Sarantitis, Apostolos Chondronasios

Abstract:

To achieve the objective of almost zero carbon energy solutions by 2050, the EU needs to accelerate the development of integrated, highly efficient and environmentally friendly solutions. In this direction, district heating and cooling (DHC) emerges as a viable and more efficient alternative to conventional, decentralized heating and cooling systems, enabling a combination of more efficient renewable and competitive energy supplies. In this paper, we develop a forecasting tool for near real-time local weather and thermal energy demand predictions for an entire DHC network. In this fashion, we are able to extend the functionality and to improve the energy efficiency of the DHC network by predicting and adjusting the heat load that is distributed from the heat generation plant to the connected buildings by the heat pipe network. Two case-studies are considered; one for Vransko, Slovenia and one for Montpellier, France. The data consists of i) local weather data, such as humidity, temperature, and precipitation, ii) weather forecast data, such as the outdoor temperature and iii) DHC operational parameters, such as the mass flow rate, supply and return temperature. The external temperature is found to be the most important energy-related variable for space conditioning, and thus it is used as an external parameter for the energy demand models. For the development of the forecasting tool, we use state-of-the-art deep neural networks and more specifically, recurrent networks with long-short-term memory cells, which are able to capture complex non-linear relations among temporal variables. Firstly, we develop models to forecast outdoor temperatures for the next 24 hours using local weather data for each case-study. Subsequently, we develop models to forecast thermal demand for the same period, taking under consideration past energy demand values as well as the predicted temperature values from the weather forecasting models. The contributions to the scientific and industrial community are three-fold, and the empirical results are highly encouraging. First, we are able to predict future thermal demand levels for the two locations under consideration with minimal errors. Second, we examine the impact of the outdoor temperature on the predictive ability of the models and how the accuracy of the energy demand forecasts decreases with the forecast horizon. Third, we extend the relevant literature with a new dataset of thermal demand and examine the performance and applicability of machine learning techniques to solve real-world problems. Overall, the solution proposed in this paper is in accordance with EU targets, providing an automated smart energy management system, decreasing human errors and reducing excessive energy production.

Keywords: machine learning, LSTMs, district heating and cooling system, thermal demand

Procedia PDF Downloads 112