Time series forecasting using ensemble learning methods for emergency prevention in hydroelectric power plants with dam

Highlights

• Faster decision-making process is necessary to prevent emergency.

• Bagging, boosting, random subspace, bagging plus random subspace, and stacked generalization ensemble models are presented.

• Ensemble models have higher accuracy than long short-term memory (LSTM).

• Support vector regression (SVR) is base learner of each ensemble strategy.

Abstract

In hydroelectric plants, the responsibility for the operation of the reservoirs typically lies with the national system operator, who controls the level of the reservoirs based on a stochastic problem for the economy of the potential energy available in the reservoir. However, in an emergency, the responsibility for the operation and control of the reservoir becomes the plant’s management. To have a faster decision-making process, it is important to have a forecast of water affluence in relation to the turbine capacity and use of the spillway. With the objective of evaluating the forecast increase in the level of the reservoir of a hydroelectric plant, this paper compares the use of the bagging, boosting, random subspace, bagging plus random subspace, and stacked generalization ensemble learning models to analyze this problem. The case study is based on data from a 690 MW hydroelectric plant, which has a 94 km reservoir and a 185 m high dam. The random subspace and stacking models had the best results for lower error, with a low time required for convergence in relation to the other models. The ensemble models resulted in greater accuracy for the assessed problem than long short-term memory.

Introduction

Based on the historical water affluence forecast for a period, the daily plant operation program (DPOP) is created, which provides the necessary requirements to meet the contractual and control needs determined by the National Electric System Operator to meet the demand required for the period [1]. The problem of forecasting affluence in relation to the generation of energy is complex [2], given that its poor planning can cause high costs for the future electricity generation or in the spillage of water. This stochastic problem is evaluated through optimizers that view energy savings over time [3].

In an emergency in which the safety limit in relation to the reservoir capacity can be exceeded, the plant controls the reservoir using the plant’s operation manual. Considering that the control of the plant is directly linked to information from external bodies [4], [5]. If there is a delay in passing on this information in an emergency, the decision-making time will become short, which may expose the plant to a safety risk. The use of information recorded based on the region’s water affluence curve in relation to the volume of rain can be mathematically modeled to predict risk situations and thus alert operators so that they are prepared for decision-making in advance.

Some researchers conducted promising studies focused on time series forecasting to assess the load on the electrical system. Kaboli et al. [6] showed that load forecasting can improve the operation of the electrical system, considering that the electrical generation must follow electrical load demand. According to Kazemzadeh et al. [7] load forecasting is one of the main studies necessary for the planning and expansion of the electrical power system, in this approach, techniques combined with optimization algorithms are a promising alternative for the evaluation in question. The uncertainty associated with the need for generation leads to problems in the management of smart active distribution networks, so an accurate forecasting model can improve planning for the operation of power systems [8].

Regarding time series forecasting, fuzzy systems are being used by some authors [9]. In [10] a fuzzy interval model based on multiple time-frequency spaces and the induced ordered weighted average aggregation operation is used, the signal is decomposed into empirical modes of ensemble and is reconstructed as several time-frequency spaces for visibility graphs. The results presented using the proposed model prove to be efficient for solving the problem, being more promising than well-consolidated models. Tanuwijaya et al. [11] used a first-order single-valued neutrosophic hesitant fuzzy time series for stock index forecasting in Indonesia and Argentina. The presented results show that the proposed method outperforms the classic fuzzy time series.

The use of techniques based on deep learning for time series forecasting has been growing considerably recently [12]. Kasburg and Stefenon [13] used the long short-term memory (LSTM) to predict the generation of photovoltaic energy, with the objective of improving the capture of the sun and thus the efficiency in photovoltaic panels. In [14] a hybrid model with stacked residual LSTMs and adaptive grouping of sub-series is applied to predict multivariate time series. The results of applying the model are superior in the effectiveness in public data sets when compared to other algorithms.

With the need to predict failures in the electrical power system to keep the system working, the ensemble learning model has promising results for assessing chaotic time series [15]. The ensemble learning model is based on the divide-to-conquer paradigm, devoted to improving the models’ accuracy in classification [16], clustering, or regression tasks [17], [18]. Several weak or base models individually perform a specific task, and when its results are combined, an efficient model in terms of accuracy is obtained [19].

Specifically for flood forecasting in hydroelectric power plants, Davydov et al. [20] assessed extreme water discharges from flood control facilities. Based on this assessment, it is possible to define modes of operation for a hydroelectric plant and flood control facilities. Hydroelectric plants have influence on flood control, as depending on their operation they can quickly increase the level downstream [21], [22].

Regarding the risk related to the use of dams with large flooded areas, Sidek et al. [23] conducts a study on the impacts of a possible rupture of a dam. From this study, the flow to the downstream area is evaluated to obtain an estimate of the flooded area. According to Assahira et al. [24] and Resende et al. [25], the construction of hydroelectric plants using a reservoir has major impacts on social, ecological, and environmental as it substantially changes the hydrological regime of a region.

Davydov et al. [26] present a work to flood control based on the calculation of the maximum flow rates for specific rivers. Using this approach it is possible to manage floods from the water reservoirs of the hydroelectric power station. Chusov et al. [27] developed a study about land protection from floods based on the use of reservoirs distributed as part of a hydrosystem with a hydroelectric power plant. This research shows the importance of the management of the water flow in hydroelectric reservoirs.

Applying the LSTM model that will be evaluated comparatively in this paper, Le et al. [28] perform flood forecasting according to daily discharge and rainfall. Accuracy can reach 99 % depending on the case evaluated, this shows that there is an influence on the accuracy in the prediction depending on the data set used since the same model only reaches 87 % for an evaluation of another data set. According to Razali et al. [29] to have efficient flood control it is necessary to study the inflow design to determine the dam structure.

The ensemble learning model is used for forecasting in several areas of knowledge [30], standing out for fault diagnostics and prognostics [31] as well as component failure [32]. This model can be used to assess the risks associated with establishing relationships with corporate partners to business failure prediction models [33]. Ensemble learning methods can be applied to assess the carrying capacity of reinforced concrete columns, which are a concern in structural design and also in performance evaluation procedures [34], it may also have a promising application for predicting the useful life of industrial components [35], which is an evaluation similar to the one being proposed in this paper.

Based on the characteristics of the ensemble models of having less computational effort and greater accuracy in certain applications, the bagging, boosting, random subspace, bagging plus random subspace, and stacked generalization ensemble models are evaluated in this paper. The general purpose of this application is to predict the emergency situations in hydroelectric power plants with dam, and thus, advance the necessary decision-making in a situation of risk to the dam. The forecast analysis was carried out based on the variation of 1-hour ahead, enough time to carry out preventive actions in a hydroelectric power plant.

The contributions of this paper for hydroelectric power plants are summarized in the following:

• The first contribution is related to the technical feasibility assessment of flood forecasting to the improvement of the safety of the hydroelectric plant through a model that can anticipate the decision-making of the plant management. Subsequent action results in a change in the emergency procedure, and this was not done in this research;

• The second contribution is related to the evaluation of the use of an ensemble learning model to forecast the inflows of floods. In this respect, the well-known bagging, boosting, bagging plus random subspace, random subspace, and stacking ensembles are combined with support vector regression to forecast the inflows of floods with high accuracy. Comparative applications of different ensemble models are rare, considering that specific models for specific solutions are typically explored;

• The third contribution is about the model’s ability to process raw data, which has nonlinearities. Even with raw data the model has higher accuracy, and less error that the ensemble models can have compared to well-established LSTM models. This shows the deep learning techniques that are being used widely today may have inferior results to the ensemble models depending on the application.

Table 1 defines the symbol used in this paper.

The remainder of this paper is organized as follows: In Section 2, the operation of the hydroelectric plant is detailed and the issue of water inflow is discussed. In Section 3, the proposed method is presented. In Section 4, the results obtained are presented and discussed. Finally, the conclusions are discussed in Section 5, about the applicability of the model.