Case Study: STE-Neural Energy Production Forecast for Run-Of-River Hydro Power Plants. By Mario Arquilla and Fabio Pasut, S.T.E.

Size: px
Start display at page:

Download "Case Study: STE-Neural Energy Production Forecast for Run-Of-River Hydro Power Plants. By Mario Arquilla and Fabio Pasut, S.T.E."

Transcription

1 Case Study: STE-Neural Energy Production Forecast for Run-Of-River Hydro Power Plants By Mario Arquilla and Fabio Pasut, S.T.E. Energy, Italy ABSTRACT The possibility to accurately predict the actual tendencies in power production of renewable energy power plants, brings benefits to both the plants management and the investors. The effects of climate change and the inherent variability of the flow in streams where the power plants are installed emphasize the advantages of an accurate prediction. To this aim, a forecasting tool based on Artificial Neural Networks has been developed. The algorithm considers geographical and climatic factors as well as operational factors such as failure rates and maintenance operations. A recursive method has been applied in order to increase the forecast accuracy; input data for the self-learning routine of the neural network are continuously updated during the operation of the power plant. Different simulations have been carried out in order to test the proposed algorithm. Two case studies of practical applications in run-of-river hydro power plants, demonstrating the potentialities of the proposed forecasting tool, are presented in the paper. Direct applications ranging from an efficient planning of maintenance operations to the establishment of accurate production budgets. Introduction The theme of efficiency improvement in power plants management, with the primary aim of maximizing their production, is taking today a leading role in the hydroelectric field. People involved in the activities of Operation and Maintenance (O&M) are increasingly supported by IT tools able to plan, monitor, collect data, diagnose, etc. In this context, STE Energy has developed an energy production forecasting tool for run-of-river hydro power plants that can be integrated within monitoring and supervision platforms. The project consisted of the following main steps: - Collection, analysis and processing of operation historical data of some run-of-river hydro power plants currently in operation; - Collection of weather data; - Development of the algorithm for the energy production forecast; - Integration of the new forecasting tool within existing monitoring and supervision platforms. Data collection Regarding the collection of data relating to the operation of power plants, the functionalities of the two softwares STE-Monitor and STE-Guardian have exploited. STE-Monitor is a web-based software designed for the remote control of production facilities and it uses the OPC protocol to ensure a wide compatibility with the control systems installed on-site. Figure 1 shows the graphical user interface of the software. STE-Monitor continuously monitors 1

2 all the main operating parameters of the power plant and, for the aim of this study, the following data have been considered: - Energy production profiles; - Alarms and faults lists. Figure 1. Graphical user interface of STE-Monitor STE-Guardian is a web-based software developed in order to easily plan the maintenance activities, to coordinate the maintenance personnel and to automatically send reports to the customer. Figure 2 shows the graphical user interface of the software. The following data have been considered: - Cataloging of plant shutdowns; - Determination of the duration of plant shutdowns. 2

3 Figure 2. Graphical user interface of STE-Guardian With regard to weather data, since the forecasting tool has been tested on Italian power plants, the web application provided by the Italian Regional Agency for Environmental Prevention and Protection has been exploited. The Figure 3 shows a typical rainfall profile, measured in a weather station during several years. Rainfall data of all the available weather stations within the river basin of the considered hydro power plants have been taken into account. Figure 3. Example of a rainfall profile 3

4 Development of the forecasting algorithm The energy production of a run-of-river power plant obviously depends on the amount of precipitation and thus on the available flow rate, but not only. There are other factors, such as the programmed plant shutdowns, the internal faults, the electrical grid disturbances, the floods, the extreme weather conditions, etc., that significantly affect the energy production. For these reasons, the collection of power plant operation data is not enough; data must be opportunely analyzed, reprocessed and understood in order to make them useful. As an example, Figure 4 shows the production trend of a small run-of-river hydroelectric power compared with atmospheric rainfall; generally, the production follows the trend of precipitation, but there are cases in which other factors sensibly influence the production profile. Therefore, in practical terms, it comes to solving a non-linear problem. Figure 4. Production trend in comparison with rainfall profile For the solution of the non-linear problem, the potential of Artificial Neural Networks (ANNs) has been exploited. An algorithm in Matlab Simulink has been developed; it receives as input historical data relating to the operation of the power plant together with meteorological data and gives as output the energy production profile. Briefly, a neural network is formed by several layers: an input layer, several hidden layers composed of several neurons and an output layer. During the so-called "learning phase" of the ANN, the weights associated to the various neurons of the hidden layers are iteratively changed, in order to obtain an output as close as possible to the real data used for comparison. Figure 5 shows a schematic diagram of an Artificial Neural Network. Figure 5. Schematic diagram of an Artificial Neural Network 4

5 For the learning phase, a back-propagation approach with decreasing gradient has been adopted; it tends to minimize the deviation between the expected output and simulation results by adapting the "weights" associated with neurons in the network. The input of a neuron j-ith in the network is given by: where: Neuron = w x + θ (1) - x i are the outputs of the neurons in the previous layer; - w ij is the synaptic weight between neuron i and neuron j; - θ j is called bias; it is usually a vector of constants used to activate the activation function and it varies depending on the method used for the estimation of the output. The algorithm, used to update the coefficients of the network, is called back-propagation algorithm for the fact that the offset recorded in correspondence of a certain data is propagated backwards in the network in order to obtain the updating formulas of network coefficients. The output error is calculated with the MSE (Mean Square Error) approach: MSE = 1 2 [Target Output ] (2) During the learning phase the weights are modified depending on their contribution in accordance with the following formula: where: Δw = η MSE w (3) - Δw ij is the element i-j of the matrix of weights; - η is called learning rate and determines how the weights should be modified. Case Studies The forecast algorithm has been tested on two Italian run-of-river hydro power plants with the characteristics listed in Table 1. Table 1. main features of the considered HPPs Genivolta and Cassano HPPs Turbine Type Bulb Turbine Rated Power [kw] 1031 Rated discharge [m3/s] 26 Head [m] Figure 6. View of Cassano HPP 5

6 Two scenarios have been simulated. A production forecast over the medium-term (annual) with monthly definition has been performed for Genivolta HPP, while, for Cassano HPP, a short-term forecast (monthly) with daily definition has been considered. Scenario 1 - Medium-term forecast for Genivolta HPP For the learning phase of the ANN, two years of historical data (both operation and meteorological data) have been used; then it has been forecasted the production trend over one year, giving as input to the network only the annual rainfall profile and the programmed plant shutdowns. The results, Figure 7, show that there are significant divergences between the prediction and the actual production during the year. Overall, however, an estimation of annual energy production with an error of less than 1% has been obtained; this result is strongly dependent on the accuracy of rainfall forecast. Figure 7. Scenario 1 Comparison between real and estimated energy production profile Scenario 2 - Short-term forecast for Cassano HPP For the learning phase of the ANN, a year and a half of historical data (both operation and meteorological data) have been used; then it has been forecasted the production trend over one month, giving as input to the network only the monthly rainfall profile and the programmed plant shutdowns during the month. The results, Figure 8, show that there are small divergences between the prediction and the actual production. The maximum daily error on the average power is 24% and the maximum monthly error on the energy production estimation is 6%; also in this case it must be noted that results strongly depend on the accuracy of rainfall forecast. Figure 8. Scenario 2 Comparison between real and estimated energy production profile 6

7 Integration into the existing monitoring and supervision system The forecast tool, called STE-Neural, has been integrated into the software tools used by STE Energy for the supervision of the production facilities, as it can be seen in Figure 9 and Figure 10. Figure 9. Learning phase of the ANN As it can be seen in Figure 9, input data for the learning phase of the ANN are continuously updated during the operation of the power plant. Figure 10. Forecasting phase of the ANN Conclusions In the present paper, the development of a new energy production forecasting tool for small runof-river hydro power plants has been presented. Simulations and tests on practical applications demonstrate the potentialities of the proposed forecasting tool. STE-Neural, integrated within the platforms for planning, monitoring and controlling the production facilities, proves to be a tool that brings an important added value to the O&M services. Understanding the behavior of a generating power plant is of primary importance; it allows to properly plan the ordinary operation and maintenance activities and it helps to opportunely manage emergencies and extraordinary interventions. 7

8 References [1] A. Mazzetti, Reti Neurali Artificiali, una nuova sfida tecnologica, Apogeo, [2] S. Patarnello, Le reti neuronali, semplificare la complessità con l aiuto dell informatica, Collana informatica domani in collaborazione con ibm semea, francoangeli, [3] D. Çinar, G. Kayakutlu, Forecasting production of renewable energy using cognitive mapping and artificial neural networks, 19 International Conference on Production Research. [4] J. Andrew Howe, PhD, Simulation and forecasting of Hydrological Power Generation: An Alternative Approach, Palisade EMEA Risk Conference London, [5] T.Olason, D. Hurdowar-Castro, J. Huysentruyt, P. Kirshen, Inflow Forecasting For Short- Term Scheduling, ASCE WaterPower 97, August [6] T. Stokelj, D. Paravan, R. Golob, Short and mid-term hydro power plant reservoir inflow forecasting, IEEE, The authors M. Arquilla (S.T.E. Energy President and Chairman of the Board): born in 1966 in Belluno (Italy) he graduated in the University of Padua in Electrical Engineering. He has almost twenty years of experience in the hydroelectric sector. He is in Board of Directors of several foreign branches of STE Energy. He is contract Professor at the University of Padua in the Faculty of Electrical Engineering. He is in the board of many national and international industrial associations. address: m.arquilla@ste-energy.com F. Pasut (S.T.E. Energy Head of Innovation and Development Department): born in 1986 in Pordenone (Italy), he graduated in Electrical Engineering in the University of Padua. He has been working in S.T.E. Energy since 2010 and he collaborates with University of Padua. He designed and managed the construction of hydroelectric power plants in Italy and abroad. Currently he is involved in several projects of research & development in the fields of renewable energy technologies and smart grids. address: f.pasut@ste-energy.com 8