1. Introduction
Electric vehicles (EVs), as an alternative to fuel vehicles (FVs), are considered zero emission cars. This statement is actually a half-truth; in fact, while EVs drastically reduce polluting emissions, more EVs will require more and more energy and the question becomes: how will this energy be produced? The European Environment Agency (EEA) [
1] has tried to make assumptions about the future number of EVs and what their impact could be. In an optimistic scenario, the percentage of EVs by 2050 will be 80%; however, in a more realistic scenario, EVs in 2050 will represent half of the vehicles in circulation. It is only a hypothesis, but in future, EVs will be widespread and governments, even if the times remain unknown, will have to face the following problem: the FVs will be replaced by EVs and the electricity demand will increase very fast.
The future road transport and electric power sectors will become more closely linked if there is a wide uptake of electric vehicles in the European Union (EU). Recent findings from the EEA show that, if a hypothetical 80% of cars in 2050 will be electric, an additional 150 GW of additional electricity generation capacity would be needed across the EU. The risk, however, is that despite the great advance in mobility, global emissions will not drop and pollution will double. If most world governments, in fact, will keep relying on fossil fuel, the problem of pollution will shift from cars to power plants. The European Environment Agency suggests covering the additional 150 GW required by the advanced scenario for powering EVs with renewable energies with the following subdivision: 87 GW from wind power, 45 GW from photovoltaic, 24 GW from hydropower, and 13 GW from biomass [
1].
The last statement highlights a very interesting question in terms of water resources because the hydropower generation, which is probably the oldest renewable energy source [
2], is nowadays extensively exploited and the energy produced is already sent into the electric grid. Hydropower is already a mature technology in Europe, with an estimated total installed capacity of 294 GW [
3]. Most of the hydropower potential is developed in northern Europe and alpine regions; in particular, Italy already has an installed hydropower capacity of 21.9 GW; however, most of the unutilized potential is concentrated in eastern Europe. Today the majority of investments in Europe are concerned with the renovation of existing facilities to minimize environmental impacts, improve efficiency, flexibility, and system resilience.
However, if on one hand the large- and medium-hydroelectric sector has been exploited extensively, and at the same time, the potential reduction of water resources due to extensive agriculture and drinking water demand grows, on the other hand small-, mini-, and micro-hydropower may represent an optimal and useful solution to produce energy from small water streams supplying rural centers and hilly areas. An interesting opportunity is represented nowadays by the energy recovery from water distribution networks (WDNs) converting water pressure in excess into electric energy or using hydraulic jumps in open channels.
Pressure control is one of the most important issues in optimizing the operation of WDNs [
4]. Currently, a common practice that aims to control and reduce leakages involves the installation of pressure reduction valves (PRVs) in those points of the WDN where the pressures are high. A novel approach consists of installing pumps-as-turbines (PATs) or replacing PRVs already installed into the WDN and producing significant amount of energy converting water pressure into electric energy. Undoubtedly, the flow rate and water pressure are extremely variable at every point of a WDN; however, there are nodes of the network whose elevation is the lowest of the surrounding nodes, where often the pressure is greater than necessary. Technical literature summarized several examples about PAT devices in a real aqueduct or WDN: Muhammetoglu et al. [
5] reported the results of a PAT system installed on a by-pass line of the Antalia (Turkey) WDN and demonstrated how the system is able to produce energy between 0.7 and 7 kWh. Rossi et al. [
6] selected the optimal PAT to insert into the aqueduct of Merano, a town located in South-Tyrol. Puleo et al. [
7] analyzed the potential energy recovery obtained through the use of PAT devices installed in a district metered area in Palermo (Italy) with private tanks and their filling/emptying process. Samora et al. [
8] presented a feasibility study of the installation of micro-turbines in the WDN of Fribourg (Switzerland). Furthermore, Balacco et al. [
9] pointed their attention to coupling a PAT system for the recovery of excess pressure levels in a WDN throughout the day with a dedicated system that supplies electric energy for the recharging points of EVs with the double advantage of turbinating water energy otherwise lost and providing electric energy “on site,” avoiding saturation of electric grids by putting energy into the same ones.
Nowadays, PATs are an optimal compromise for the growth of energy production in a WDN thanks to their low cost compared with traditional turbines, a wide range of sizes and specific speed numbers. However, even if the technical literature is rich in studies, models, and pilot experiences [
2,
10], water management authorities and local administrations still fear and doubt PATs’ effectiveness for the recovery of energy into a WDN. Many doubts are due to scarce technical information provided by pump manufacturers about their machines operating under reverse mode. In addition to the above-mentioned limits, it should be highlighted that there is a large variability of hydraulic working conditions in a WDN and the necessity to guarantee pressure head at each node within a well-defined interval, which is indispensable to satisfy water demand at any time. Considering these aspects, Carravetta et al. [
11] proposed three different approaches to regulate a PAT inserted in parallel into a WDN: a hydraulic regulation thanks to control valves, an electric regulation thanks to inverters, and a combined regulation system.
The best efficiency point (BEP) parameters in turbine mode (
BEPt) are quite different from those in pump mode (
BEPp); the relation between the pump and turbine modes is not the same for all type and size of machine and depends on the specific speed and the losses incurred expressed in terms of machine efficiency [
12]. For the sake of clarity, flow through impellers of a machine in a pump mode is subjected to a secondary flow known as “circulation loss,” which is due to the rotation of the impeller; this secondary flow in reverse mode is negligible since it is located at the inner periphery of the impeller. Because of this phenomenon, both the flow rate and head at the BEP are bigger in reverse mode than in normal operation.
Based on this consideration, a correction factor is normally adopted for both flow rate and head rate:
A different approach has been adopted by a few researchers [
13,
14,
15,
16] by representing numerical and experimental results with a non-dimensional approach:
and defining several regression curves that put these three parameters in relation to each other. In particular, Fontana et al. [
17] conducted an experimental and numerical analysis and highlighted how the regression curves for
ψ and
π can be better described with the following equations:
These equations showed a very good agreement with experimental data inferred for a flow number < 30 since they observed how the optimal operation of a PAT is reached for a value of about 0.12, and for this reason, regression curves can be obtained by fitting data for a restricted flow number field.
As a whole, the common interest of part of the scientific community is still to define a general methodology for pump performance in a reverse mode, known as the optimal conditions of operation of a pump.
Table 1 summarizes some of suggested models to predict PAT performance: flow ratio (
q) and head ratio (
h). Expressions have been subdivided in chronological order and grouped into two classes, the first based on the BEP parameters [
18,
19,
20,
21,
22,
23,
24,
25,
26,
27] and the second based on the specific speed number (
Ns) [
13,
28,
29,
30,
31,
32,
33].
The same table shows how there is still no clarity on this issue, expressions are often very different from each other in numerical terms, and the parameters on which they depend on are not always the same.
In this framework, and considering the complexity and difficulty to define a general law for the BEP of a PAT, a beneficial strategy is to adopt an estimation approach for the PAT selection that investigates the role of the candidate input parameters with the aim to define the model structure. An estimation approach permits one to seek all possible models; nevertheless, it is necessary to identify those having a physical meaning. Starting from this consideration, Venturini et al. [
34] presented a comparison of different approaches that can be used for the prediction of a PAT: a physics-based model; two models called “gray,” supported by literature data and based on a priori knowledge of the physical process but without a priori knowledge of the parameters influencing the process; and finally an evolutionary polynomial regression (EPR) model using field data. Values predicted by the four models and compared to the original literature experimental data [
30] showed a good agreement with the latter, even if the EPR goodness is clearly well-founded on the input database dimension. Rossi and Renzi [
35], using a non-dimensional approach, adopted artificial neural networks to forecast both BEP and PATs performance in reverse mode by fitting operating data extrapolated from technical literature. However, the obtained formulas in both applications, even if showing a very good fitness, are extremely complicated and characterized by a polynomial structure with numerous terms that are difficult to interpret from a physical point of view.
The technical literature (see
Table 1) shows a considerable dispersion in terms of formulas to predict the performance of a pump in turbine mode. The parameters that influence the phenomenon in each formula not only change, but above all, the exponent with which to represent them is different for each of them. In the light of the above-mentioned studies, the present paper conducted a sensitivity analysis on a PAT efficiency while knowing the pump efficiency in direct mode. Sensitivity analysis has been carried out by comparing results of artificial neural networks (ANNs) and EPR methodologies with the aim to understand the real weight of every input parameter on the evaluation of a PAT performances while knowing those of a pump in a direct mode.
In detail, the paper makes use of a multi-objective genetic algorithm (MOGA) strategy to address the optimal design of ANN [
36,
37] with the aim of finding the optimal trade-off among parsimony and accuracy for the returned models. In addition, EPR has been applied with the aim to identify a relation between some of the input candidates and every output parameter.
In order to guarantee a more general applicability of the prediction model for PAT performance, the analysis is supported by an input literature database [
6,
13,
16,
30,
31,
33,
38,
39,
40] consisting of pumps operating in a range of specific speeds from 9 to 80 (rpm) and characterized by several impeller sizes.
The rest of this paper is organized as follows.
Section 2 of this paper summarizes the adopted methodology,
Section 3 describes the considered case study, and finally
Section 4 presents the obtained results and a comparison between two adopted methodologies.
5. Conclusions
The paper conducted a sensitivity analysis on the input parameters that most influence a pump performance in reverse mode and offers a comparison between ANNs and EPR methodologies in order to increase knowledge about different models for the prediction of PAT behavior. Literature data summarized several studies based on experimental tests or numerical application, and overall, these methods can be subdivided into a first category based on the best efficiency point (BEP) parameters and a second one based on the specific speed number Nsp. Starting from this knowledge, the paper provides a numerical analysis with the aim to understand the real weight of every input parameter on the output ones.
The ANN approach, while containing the intrinsic limit of not providing a well-defined formula, gives the user the possibility to identify the structure of the best model, emphasizing the correlation among the most influential input factors and the output [
45]. The EPR technique, instead, is able to provide a clear formula, but in this study, this research was not simple and understandable in all cases. The use of one methodology did not exclude the other because while ANN needs a preliminary selection of the transfer function, the latter does not need an a priori knowledge of the model structure.
Overall, the paper, thanks to the application of ANNs, highlighted the real weight of each input parameter on the outputs, and this is certainly an advantage and benefit to the scientific research on this topic in the near future. Whereas, EPR application returned general formulas with a high coefficient of determination values, the selected formulas for every output parameter had values of CoD of about 0.75, except for the specific speed Nst, where the value was equal to 0.95.
Both methods were not able to relate the flow ratio
q and head ratio
h to the input candidate variables, in agreement with Reference [
35], and this should mean that it is probably necessary to deepen this aspect with wider experimental tests.
The results of this sensitivity analysis point out that while parameters like Nst certainly depend on well-defined input parameters, others like q and h instead show a low correlation with the input parameters, and for this reason, they are probably not able to provide a general formula for predicting the pump performance in a reverse mode, knowing those of a pump in a direct mode. As a whole, the study showed how the application of two different methodologies from a mathematical point of view leads to absolutely consistent results.
In conclusion, the results of this study may represent a starting point for future research, thanks to experimental apparatus and numerical approaches, that will be able to make use of this sensitivity analysis and to focus its efforts on the most influencing input parameters, such as the specific speed Nst, on the evaluation of PAT performances.