Management of inflow forecasting studies

Brazilian power generation is predominantly hydroelectric. It has several characteristics: a high number of plants operating jointly in the same watershed, hydrologic diversity of the regions, large distances between the natural resources and the main consumption areas, wide transmission network, and the presence of multiple economic agents with different interests. According to the Electric Energy National Agency, despite the incentive to use alternative energy sources, hydroelectric power will remain the main source of electricity in the country for many years (ANEEL 2008).

In this context, an important activity for water resources management is inflow forecasting, which can be used in decision support systems, reservoir operations, cascade coordination, and the determination of energy sale prices.

Inflow forecasting is one of the most difficult tasks in hydrology. It is characteristically dynamic, uncertain, and nonlinear. However, it is important to have accurate estimates of the variables involved in hydroelectric planning so that software, used for optimization and simulation of the system, provide reliable results (Lopes 2007; Hidalgo et al. 2010, 2013; ONS 2014).

In general, inflow forecast models are divided into three categories: deterministic, conceptual, or parametric. Deterministic models describe the rainfall–runoff process using physical laws of mass and energy transfer. Conceptual models provide simplified representations of key hydrologic processes using a perceived system. Parametric models use mathematical functions to relate meteorological variables to inflow (Dawson & Wilby 2001; Zhang et al. 2009).

Several research projects have been conducted to formulate inflow forecasting models. Souza Filho & Lall (2004) apply linear regression; Dawson & Wilby (2001) and Gomes & Montenegro (2010) use artificial neural networks; Bravo et al. (2009) employ a multilayer, feed-forward, artificial neural network and a distributed hydrologic model. Coulibaly et al. (2005) combine nearest-neighbor, conceptual, and artificial neural network models. Fadiga Jr et al. (2008) analyze hydrologic and stochastic models, and a third model, which results in a linear combination of the first two.

Inflow forecast models have specific characteristics, such as type and number of parameters and input data, and handle a considerable amount of data. It is often hard to quickly compare them to improve the accuracy of prediction of water inflow. Some systems can run inflow forecasting, such as, PREVIVAZ (Maceira et al. 1999), PREVIVAZH (Costa et al. 2000; Livino et al. 2001), GEVAZP (Jardim et al. 2001), PREVIVAZM (Costa et al. 2003), and CPINS (Acioli et al. 2004). However, generally, they cannot manage studies from different models.

This paper proposes a standard structure that will enable distinct inflow forecast models to be run in a single platform. In this approach, the models share the same database and interface to input and output data, facilitating the management and analysis of different inflow forecasting studies and models.

Inflow forecasting involves a set of plants or monitoring stations, input data, model parameters, and output data. A series of prediction studies can be conducted for each model. To manage the studies, the proposed method is divided in four parts: common data for the models, specific data for the studies, interface, and advanced assessment module, as Figure 1. It can be applied to models with parameters previously calibrated to the same set of plants or stations, and to the same period of data.

The first part of the method, ‘Database’, is related to the common data manipulated by the models, such as data identifying the plant or station (ID, name, acronym, etc.) and variable input data (observed inflow, observed and predicted rainfall, etc.). These must be recorded in a database and shared among the models. The database must be complete enough to cover the requirements of all models that make up the structure.

The second part, ‘Text Files’, is associated with the specific data of the prediction studies, which must be saved in a set of text files. Each study must have identification, parameters, data, and results files. The identification file contains data characterizing the study. They will be entered automatically by the system, although some could be customized by the researcher-user – e.g., the study name and its start date. The parameters file holds the coefficients used in the prediction calculations, which will be embedded in the system but could be changed by the researcher-user. The data file must contain the input data for the predictor in use, as extracted from the database, although they could also be altered by the researcher-user. The results file will contain the output data from the model in use, i.e. the predicted inflows. If the study is applied to a previous period, the observed inflows must also be saved here.

The third part of the method, the ‘User Interface’, concerns the interface for the researcher-user. It shows the available prediction models, and the set of plants and stations for which the models can be applied. It also provides means to create, save, rename, delete, and/or run an inflow forecasting study. The four text files associated with a study must be accessible to the researcher-user via the interface.

The fourth part of the method involves the ‘Advanced Assessment’ module, comprising three sub-modules – queries builder, statistical analysis, and graphical analysis.

The queries builder is important for analyzing the input data for the models. According to Bravo et al. (2009), the performance of the predictors depends strongly on the quality of the precipitation forecasts. In general, it is necessary to know the database, the relationship between the tables, and the structure query language (SQL), to build questions (Ben-Gan 2012). The proposed method suggests automatic construction of the SQL command responsible for extracting information from a database. In this way, the researcher-user can run queries without the need for specific technical knowledge of the computation area.

For statistical analysis, it is recommended that the numerical results of the studies are shown on a grid, which should present the maximum, minimum, average, and median predicted values. It can also exhibit the performance metrics frequently used for validation of hydrologic models, such as mean absolute percentage error, root mean square error, mean absolute error, and mass curve coefficient (E).

From the graphical analysis, the researcher-user must be able to view the trajectory of the predicted inflows on a graph. If the study applies to a previous period, the trajectory of the observed inflows must also be presented. If the same study is run by more than one model, the results of all of them must be shown in a single window. This makes visual comparison of the models easier. It is also useful to present the precipitation graph close to the inflow graph, as this facilitates evaluation of the forecast results as a function of rainfall in the period.

The proposed method has been applied to five mathematical models based on linear regression, artificial neural networks, and hydrologic simulation. The two models (daily and monthly) using linear regression were developed by researchers from Civil and Environmental Engineering, State University of Campinas, in 2009. The two models (daily and monthly) using artificial neural networks were developed by researchers at the Institute of Economy at the same university in 2007 (Luna 2007; Luna & Ballini 2011). The daily, hydrologic simulation model is based on research at the Electric Energy and Water Department of Sao Paulo (Lopes et al. 1982; Lopes & Porto 1993).

The daily forecasts are performed 15 days ahead, while the monthly forecasts cover the period 6 months ahead. The models were calibrated for a set of 10 hydroelectric plants and 15 monitoring stations using data from 2000 to 2011. The plant and observed inflow data were provided by the company that manages the plants, AES Tietê (AES 2013). The rainfall data were obtained from Centro de Previsão de Tempo e Estudos Climáticos (CPTEC 2013). The plants and stations are located on the rivers Rio Grande, Tiete, and Pardo. To implement the proposed structure, the C ++ programing language (Stroustrup 2013) and the SQL Server (Ben-Gan 2012) were used.

For the first part of the method, the common data manipulated by the five models were identified and recorded in three database tables: ‘Plant Station’, ‘Daily Data’, and ‘Monthly Data’. The ‘Plant Station’ table stores the ID, name, and acronym of a plant or station. The ‘Daily Data’ table stores the fields: plant or station ID, date of record, inflow, rainfall, and type of rainfall. The ‘Monthly Data’ table contains the same fields as the previous table but stores the respective monthly data.

The specific data for the studies were organized in four files: identification, parameters, data, and results. The second and third files (parameters and data) are adjusted according to the model with which the study is associated. The identification file records: creation date, category (daily or monthly), model type (used method), study name, start date, number of forecast intervals, and comments. The parameters file holds the equation coefficients. The data and results files cover the input and output data for the study, respectively. All except the results file can be changed by the researcher-user.

The interface was set up to manage the prediction studies using panels and tree views. In the panels, the studies are organized in two groups: daily and monthly forecasts. Within each panel, tree views display the models (as parent nodes) and their studies (as child nodes), as shown in Figure 2(a). The set of plants and stations to which the models can be applied was organized into a list in which plants and stations have specific identification icons – see Figure 2(b). The top three items are plants, the others stations. The four text files associated with a study were organized in an interface sheet. Examples of the files can be seen in the next section.

The fourth part of the proposed method concerns the queries builder, and statistical and graphical analysis. The next section shows the grid and graphs created for the last two. For the queries builder, the researcher-user must choose the fields to be examined, the filter conditions, and the ordering of results (ascending or descending), as in Figure 2(c). The module then creates the specific SQL command automatically to extract the information from the database.

The benefits of the proposed method are analyzed in four scenarios, as presented in Table I. In the first, the suggested structure is evaluated for water inflow prediction. In the second scenario, it is used to analyze the performance of a specific model. In the third scenario, it is explored as a research tool for inflow forecasting. In the fourth, it is employed as a comparison tool among the models.

In the first scenario, two actions must be executed to predict inflows. The first is to select the panel with the intended prediction (daily or monthly) and the position in the tree view corresponding to the desired model (linear regression, artificial neural networks, or hydrologic simulation). The second is to save the settings and run the prediction study. The results can be seen on a graph, a grid, or a text file.

In the second scenario, analysis of the performance of a model can only be applied in relation to a previous period, so that the predictions can be compared with the observed values. Figure 3(a) shows the graph that allows comparison between observed and predicted inflows, in this case a monthly study for Graminha plant. It covers the period from March 2000 to June 2005. Figure 3(b) shows the observed precipitation over the same period, as suggested in the Methodology section.

In the third scenario, the structure obtained by applying the proposed method can be used to improve models. In this case, the text files are used. For example, techniques can be refined by changing the models’ coefficients (parameters file), to analyze their impact on the corresponding outputs (results file). As in the second scenario, the starting date of the forecast (identification file) can be altered in order to compare predicted and observed inflows. It is also possible to change the input data (data file) to analyze the impact of that on the final result. The queries builder can also be used to aid researchers in this process. Figure 4(a) presents the parameters file with the monthly linear regression model coefficients estimated for some plants and stations. The parameter file content is specific for each model.

For the fourth scenario, it is interesting to compare the results presented from several models in a single graph, as the proposed structure can manage studies from different inflow models. The results of the studies selected must be capable of being displayed in a graph. Figure 4(b) shows a graph that facilitates comparison between model results. The selected studies predict the monthly inflows to the Graminha plant for the period December 2009–May 2010.

This paper presents a method for integrating different inflow forecasting models into a single structure. It aims to standardize the management of prediction studies and the presentation of their results, facilitating the analysis of models.

The method was applied to five numerical models calibrated using data from a set of 10 hydroelectric plants and fifteen monitoring stations. Data from 2000 to 2011 were used to calibrate the models.

The results show three main benefits from the proposed method. Firstly, separation of the ‘common data for the models’ and the ‘specific data for the studies’. The former are stored in a database, the latter saved in a set of text files. The aim is to preserve the original database information, which is shared by the different models, while allowing the researcher direct access to the information manipulated by the studies. Thus, it is possible to change the case study data (for testing) without compromising the database content. The use of database and text files also facilitates team collaboration, i.e. the exchange of case studies among researchers, because text files are generally much smaller than database files.

The second benefit is presentation of the text files within the tool. Users do not need to make external changes to the environmental system to change study file content during the tests.

The third benefit is related to the advanced assessment module. It provides a means to evaluate the database (using the queries builder), analyze prediction studies, and compare models (using grids, graphs, and text files).

In future, the authors intend to use the structure created to evaluate model performance. They will aim to answer questions such as which model performs better in relation to dry or rainy periods, which model gives better results in atypical situations, and which model has the smallest standard deviation between the predicted and observed inflows for a certain period.

In conclusion, the proposed method is a versatile tool for predicting inflows, analyzing model performance, being used as research tool, and for comparison between the outcomes from distinct models. It can be applied for any set of mathematical models developed for the same purpose.

The research reported herein was supported by AES Tietê company and FAPESP, Brazilian Government Agency dedicated to the development of science and technology.