The Value of Perfect Forecasting in Optimizing

the Management of Energy Communities

Patrizia Beraldi

, Luigi Gallo

and Alessandra Rende

Department of Mechanical, Energy and Management Engineering, University of Calabria, Italy

Keywords:

Machine Learning Techniques, Solar Production, Forecasting, Renewable Energy Communities.

Abstract:

The rise of Renewable Energy Communities (REC) is transforming energy systems by promoting decentral-

ized renewable energy production, but their operational efﬁciency is hindered by the inherent uncertainty of

production sources like photovoltaic systems. Accurate day-ahead forecasting is pivotal for optimizing REC

energy management strategies, balancing production, consumption, and grid reliance. This study evaluates

ﬁve machine learning (ML) models—Support Vector Regression, Random Forest, Adaptive Boosting, Gradi-

ent Boost Regression Tree, and Stacking Generalization—against standard accuracy metrics and introduces

the Value of Perfect Forecast, a novel metric that quantiﬁes the economic impact of forecast inaccuracies on

REC optimization. Results indicate that, while some models perform better in standard accuracy metrics,

others are more effective in reducing the economic impact of forecast errors, emphasizing the necessity of

aligning forecasting approaches with optimization goals to achieve meaningful operational improvements.

1 INTRODUCTION

In recent years, Energy Communities (ECs) have

gained increasing attention as a reference model for

driving the energy transition. Deﬁned as coopera-

tive or collective groups of local stakeholders, ECs

typically consist of individual renewable energy pro-

ducers, such as homeowners with photovoltaic (PV)

panels, as well as small and medium-sized enterprises

and public institutions. If an EC’s energy production

comes primarily from green sources, it is classiﬁed as

a Renewable Energy Community (REC).

The primary function of a REC is to facilitate the

sharing of locally produced renewable energy among

members of the community. This replaces the concept

of self-consumption, typically considered at the indi-

vidual level, with a broader concept of ’virtual’ self-

consumption. Ideally, energy needs should be met

within the community, thereby reducing dependence

on the grid, with consequent economic beneﬁts.

In such a setting, members can achieve cost sav-

ings, by, for example, receiving surplus energy pro-

duced by a neighbour at rates signiﬁcantly lower than

current electricity tariffs. On the other hand, a coali-

tion member can be incentivized to share the pro-

https://orcid.org/0000-0002-1672-4033

https://orcid.org/0009-0002-0553-4362

duced energy, receiving compensation above to net

metering or feed-in tariffs. Beyond the economic ben-

eﬁts, social and environmental motivations are also

driving the adoption and expansion of RECs. Indeed,

they contribute to reducing carbon emissions in line

with global climate targets and environmental stew-

ardship. In addition, RECs promote social cohesion

by strengthening local networks, making communi-

ties less vulnerable to energy price ﬂuctuations and

external supply disruptions. From the REC’s per-

spective, achieving a high level of self-sufﬁciency by

optimally matching production and consumption is a

challenging task due to the uncertain and intermittent

nature of renewable generation. In this evolving con-

text, accurate forecasting becomes increasingly im-

portant to support the effective operation of the REC’s

resources. Without accurate forecasts, energy system

optimization can suffer, leading to operational inefﬁ-

ciencies and increased reliance on the external grid.

These inefﬁciencies can limit the beneﬁts to coalition

members and also reduce the attractiveness of REC

membership.

In this paper, we focus on the optimal operation of

a REC with the aim of investigating the value of the

perfect forecast. The approach relies on the idea of

combining predictive and prescriptive methodologies

to deﬁne a robust tool to support decision-makers.

The predictive analysis plays the key role of reducing

Beraldi, P., Gallo, L. and Rende, A.

The Value of Perfect Forecasting in Optimizing the Management of Energy Communities.

DOI: 10.5220/0013343700003893

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Operations Research and Enterprise Systems (ICORES 2025), pages 177-185

ISBN: 978-989-758-732-0; ISSN: 2184-4372

177

the uncertainty intrinsic in RES generation, whereas

the prescriptive methodology beneﬁts from accurate

forecasts to provide more reliable solutions.

Most studies assess predictive techniques mainly

using standard performance metrics. However, these

evaluations often overlook how well these techniques

contribute to the effectiveness of solving real-world

optimization problems. To address this gap, we

introduce a new metric, called the Value of Per-

fect Forecast (VPF) which quantiﬁes the cost asso-

ciated with uncertainty in the decision-making pro-

cess. The VPF parallels the well-known Expected

Value of Perfect Information, used in stochastic pro-

gramming (Ruszczy

nski and Shapiro, 2003). The

VPF measures the additional costs incurred when op-

erational plans, determined by considering imperfect

forecasts, must be adjusted to account for actual data.

By comparing these costs to those derived from an

optimal plan based on perfect information, the VPF

provides a clear representation of the value decision-

makers might assign to eliminating forecast uncer-

tainty, thereby offering a more comprehensive evalu-

ation of forecasting techniques in practical optimiza-

tion contexts.

The idea of measuring the value of the forecasting

techniques, not only in terms of accuracy, has only

been partially explored in the scientiﬁc literature. (Pe-

terssen et al., 2024) studied the impact of forecasting

on the optimization of energy systems. Speciﬁcally,

they compare the solutions of a linear programming

problem, where the input parameters are assumed to

be perfectly known, with a priority list, i.e. a heuristic

strategy that does not use any forecasts at all. The re-

sults show that when no forecast is taken into account,

there is a cost increase of 28%, whereas the use of

limited forecasting can reduce this limit to 22%. Sim-

ilarly, (Putz et al., 2023) assess the impact of forecast

accuracy on local ECs. Their study ﬁnds that relying

solely on conventional quality metrics for selecting

a forecast approach fails to capture its true value in

supporting EC operations. In their case study, where

forecasts relate to electricity and thermal loads, the

authors show that signiﬁcant improvements in fore-

cast quality yield only marginal gains in terms of KPIs

for the EC.

Our study aims at further investigating this is-

sue, highlighting the importance of contextualizing

the forecast methods in the interplay between predic-

tive and prescriptive methods. Speciﬁcally, we fo-

cus on predicting energy production from PV panels

owned by REC members. To this end, we implement

ﬁve machine learning (ML) techniques: Support Vec-

tor Regression (SVR), Random Forest (RF), Adap-

tive Boosting (ADA), Gradient Boost Regression Tree

(GBRT), and Stacking Generalization (STK). In ad-

dition to evaluating ML techniques against traditional

KPIs, we measure their practical impact by incorpo-

rating their predictions into a prescriptive optimiza-

tion framework aimed at deﬁning the best daily op-

erating plan. The rest of paper is organized as fol-

lows: Section 2 outlines the forecasting techniques

for daily PV production. Section 3 details the opti-

mization model. Section 4 introduces the KPIs used

in the evaluation process. Section 5 describes the case

study, and Section 6 presents and analyzes the results.

Conclusions and future reseach directions are drawn

in Section 7.

2 FORECASTING METHODS

The scientiﬁc interest in the design of more and more

accurate forecasting algorithms for the electricity pro-

duction from renewable energy sources is evident by

the very large number of publications on this subject.

We refer interested readers to recent contributions that

survey the main algorithms. Among the others, we

cite (Sharadga et al., 2020), (Yao et al., 2019), (Ma

et al., 2022), (Kodaira et al., 2021).

Here, we focus on the forecasting approaches for the

day-ahead electricity production from PV panels and

we brieﬂy describe the main methods implemented in

our study.

2.1 Support Vector Regression

The ﬁrst technique analyzed in our study is the SVR.

Let us consider a training dataset, represented by the

set of pairs {(x

), ..., (x

)}, where x

∈ R

is a

vector of n input features, and y

∈ R denotes the cor-

responding PV power generation values. The origi-

nal feature space is mapped into a higher-dimensional

space through a kernel function. This mapping en-

ables the model to handle nonlinear relationships be-

tween the input features and the target variable. In

this study, we have used the radial basis function, that

transforms the input feature vector x

into a new fea-

ture representation ϕ(x

), via the transformation map

Φ (Ding et al., 2021). The primary goal of the SVR

is to estimate a function f (x) that approximates the

relation between input variables and target:

y = f (x) = ωϕ(x

) + b.

Here ω is the vector of weights associated with the

input variables, and b is the bias term. These param-

eters are determined by solving the soft margin opti-

ICORES 2025 - 14th International Conference on Operations Research and Enterprise Systems

178

mization problem, deﬁned as follows:

min

∥ω∥

∑

i=1

(ξ

+ ξ

−

) (1)

s.t.

− ωx

− b ≤ ε + ξ

i = 1,...,N (2)

ωx

+ b − y

≤ ε + ξ

−

i = 1,...,N (3)

, ξ

−

≥ 0 i = 1,...,N (4)

where ξ

, ξ

−

are slack variables that allow deviations

beyond the ε margin. The objective function balances

complexity, controlled by the ﬁrst term in (1), with

penalties (second term) for errors beyond the toler-

ance ε, regulated by the parameter C (Vapnik, 1998).

2.2 Random Forest

In this study, various ensemble methods are applied

to model the relationship between the PV power gen-

eration and input features. Ensemble learning com-

bines multiple basic learners to enhance predictive

performance compared to individual models. The ﬁrst

method employed is Random Forest with out-of-bag

validation (Breiman, 2001). RF is a bagging method,

i.e., a technique used in regression tasks that trains

multiple base learners on different subsets of the data

and combines their outputs to produce a more ro-

bust prediction. In this approach, multiple subsets

(called ”bags”) are sampled with replacement from

the training dataset. Each bag is used to train an indi-

vidual predictive model, resulting in a collection of

models. The ﬁnal prediction of PV power genera-

tion is obtained by averaging the predictions from all

models. This aggregation reduces overﬁtting and im-

proves generalization. In our experiments, the CART

algorithm (Breiman et al., 1984) has been used to train

the multiple regression trees (RTs) that serve as base

learners.

2.3 Adaptive Boosting

The other method employed in this study is the Adap-

tive Boosting (ADA). Unlike bagging methods, which

train models independently, boosting focuses on se-

quentially training models such that each subsequent

model learns from the errors of its predecessor. In our

application, ADA trains multiple RTs sequentially.

The training process begins by associating an equal

weight ω

to N data points that compose the train-

ing set. These weights are used to compute the prob-

ability for sampling data points p

∑

i=1

. Through

this probability distribution, the training set is sam-

pled with replacement. The ﬁrst RT, RT

, is trained

on the sampled data set and generates the prediction

ˆy

= RT

). A linear Loss function is computed for

each data point i:

ˆy

− y

where D = sup | ˆy

− y

|. Based on the loss values,

ADA updates the weights of the data points, increas-

ing the weights of those with higher errors. This ad-

justment ensures that subsequent learners focus more

on data points that the previous learner struggled

with. The procedure for weight updates and assign-

ing importance to each regression tree is detailed in

(Drucker, 1997). The ﬁnal hourly PV generation fore-

cast is computed as the weighted median of the pre-

dictions from all the RTs in the ensemble.

2.4 Gradient Boosting

Another technique used in our study is GBRT, where

the RTs are iteratively trained to minimize the pre-

diction error (Friedman, 2001). The algorithm begins

with an initial RT, RT

which approximates the rela-

tionship between the target variable y and the input

feature vector x. The initial prediction for each data

point i are ˆy

= RT

). At step t = 1, the next RT

trained to learn the residuals r

deﬁned as the gradient

of the loss function

δL(y, ˆy

)

δ ˆy

with respect to the previous prediction. The predic-

tions are then updated by adding the contribution of

, scaled by a learning rate η:

ˆy

= ˆy

+ ηRT

This process continues iteratively, At each step t, the

tree RT

is trained to minimize the residual errors of

the current predictions, reﬁning the model incremen-

tally. The ﬁnal prediction y

is given by

ˆy

= RT

) + η

T −1

∑

t=0

In our study, the Mean Squared Error has been used

as loss function L, minimizing the average squared

differences between predicted and true values. The

learning rate η is an hyper-parameter tuned during the

training process with ten-fold cross validation.

2.5 Stacked Generalization

As an extension to the previously discussed meth-

ods, this work also incorporates a stacked generaliza-

tion (STK) (Wolpert, 1992). Stacking combines the

The Value of Perfect Forecasting in Optimizing the Management of Energy Communities

179

strengths of multiple base learners by using their pre-

dictions as inputs to a higher-level model, called the

meta-learner or blender. The aim is to build a robust

forecasting method by adjusting the results of the sub-

models and minimize the ﬁnal prediction errors. Let

us denote by J the set of tested ML models and let ˆy

be the corresponding forecast. Then, the ﬁnal predic-

tion is deﬁned as:

ˆy =

∑

j∈J

ˆy

To determine the coefﬁcient β

a linear regression

model has been used, by minimizing the error be-

tween the stacked ensemble’s predictions ( ˆy

) and the

actual observed values. By ﬁtting β

, the meta-learner

assigns greater weights to models with better predic-

tive performance.

3 OPTIMIZING THE REC

MANAGEMENT

With the aim of evaluating the impact of perfect

forecast on energy systems, we consider the prob-

lem of deﬁning the optimal daily operational plan of

a REC, where some members own generation units

(prosumers), whereas others are simple consumers.

The centralized management of the REC entails the

deﬁnition of a shared strategy, where the energy re-

quests and production proﬁles of all REC’s members

are considered collectively. Compared to an individ-

ual approach, where each REC’s member optimizes

his own resources independently, the uniﬁed one ac-

counts for intra-community energy exchanges, thus

increasing the self-sufﬁciency rate with a consequent

maximization of the gain for trade. Electricity short-

age or excess production (not used within the REC),

are compensated by transactions with the power grid.

We assume that the REC is managed by an aggrega-

tor that represents the interface with the power mar-

ket and has to guarantee the demand satisfaction for

all the community’s members. The ﬁnal aim is to

deﬁne the most convenient operational plan. In our

model, we assume that the daily demand proﬁles of

the REC’s members are known in advance, whereas

the uncertain electricity production from PVs is re-

placed by its forecast. Since the accuracy of the sup-

ply proﬁle directly inﬂuences the entire operational

plan, deviations from the actual electricity produc-

tion could signiﬁcantly impact the overall costs, i.e.,

the aggregator might have to resort to the real-time

market where rates are typically less convenient com-

pared to the tariffs of the day-ahead electricity market.

To formally deﬁne the problem, we introduce the

sets N and T associated with the REC members and

the operational time horizon, respectively. In the ex-

periments, we have considered a daily horizon, with

hourly time steps. Forecasts are generated one day

in advance and serve as input data in the optimiza-

tion phase. The process is repeated according to a

rolling horizon scheme, where each day updated fore-

casts and new input data are used in the optimization

phase. Prosumers within the REC are supposed to be

equipped with battery energy storage (BES) devices.

For each n ∈ N , we denote by C

the battery size

and by D

the electricity demand at time t. As for

the energy production, we denote by G

the energy

produced by the member n at time t. Energy pro-

duced and not directly used can be stored in the BES,

if available, and used later or can, eventually, be ex-

ported to the REC. For each member n ∈ N and time

t ∈ T , the following decision variables are introduced

in the formulation:

• SoC

state of charge of the BES;

• E

, E

amount charged in and discharged from

BES, respectively;

• EC

, EC

out

energy amount imported from and

exported to the REC, respectively;

• EG

, EG

out

energy amount imported from and

exported to the power grid, respectively.

The aim is to deﬁne the operational plan that min-

imizes the total costs:

min f =

∑

t∈T

∑

n∈N

+ P

− R

out

− R

out

). (5)

Here P

and P

denote the electricity tariffs to pur-

chase electricity from the power market and the com-

munity, respectively, whereas R

and R

are the rev-

enues obtained when selling electricity to the main

grid and to the community. We assume that such

data are known in advance and that P

< P

and

> R

to encourage the sharing of electricity within

the REC.

Below we report the main constraints introduced

in our formulation. The ﬁrst set of constraints (6)

ensures the satisfaction of the load demand for each

member n of the REC and for each time period t.

Speciﬁcally, the demand can be satisﬁed by using en-

ergy imported from the grid and/or from the REC

and, eventually, by using stored energy and the self-

production. The excess can be charged to the BES

and/or sold to the grid or the community.

+ EG

+ E

+ G

+ EC

out

+ EG

out

+ E

∀t ∈ T , ∀n ∈ N

(6)

ICORES 2025 - 14th International Conference on Operations Research and Enterprise Systems

180

By (7), energy balance within the REC is guaranteed:

∑

n∈N

(EC

− EC

out

) = 0 ∀t ∈ T (7)

Constraints (8) limit the total energy that can be im-

ported to a given upper bound limit E

max

, computed

on the basis of the operation power:

+ EG

≤ E

max

∀n ∈ N , ∀t ∈ T (8)

Constraints (9)-(12) refer to the management of the

storage device:

SoC

= α SoC

nt−1

+ η

−

∀n ∈ N ,∀t ∈ T

(9)

Speciﬁcally, constraints (9) model the dynamics of

the BES, linking the state of charge of a time t, to the

amount in the battery at the end of the previous time

and to the charged and the discharged amounts. Here

α, η

and η

accounts for loss. We note that for the

ﬁrst period of the time horizon, the amount initially in

the device is set equal to the energy in the battery at

the last period of the previous day. Constraints (10)-

(12) limit the stored and the charged and discharged

amount to a percentage β of the BES capacity.

SoC

≤ C

∀n ∈ N , ∀t ∈ T (10)

≤ β C

∀n ∈ N , ∀t ∈ T (11)

≤ β C

∀n ∈ N , ∀t ∈ T (12)

Finally, the last constraints refer to the nature of the

decision variables, for every time period t and mem-

ber n:

, EC

out

, EG

,EG

out

, SOC

≥ 0.

(13)

4 ASSESSING THE FORECAST

QUALITY

We assess the quality of the forecasting methods us-

ing both traditional metrics and the new index. While

the former are inherent in any forecast, regardless of

the context in which they are used, the latter measure

accounts for the value of prediction and evaluates the

beneﬁt of incorporating the forecast into the decision-

making process.

Accuracy of the forecasting methods has been

measured by traditional KPIs. In particular, in

the experimental phase we have used the follow-

ing traditional metrics, Root Mean Squared Errors

(RMSE), Mean Absolute Error (MAE), and R-Square

coefﬁcient(R

), deﬁned as follows:

RMSE =

∑

i=1

− ˆy

)

MAE =

∑

i=1



− ˆy



= 1 −

∑

i=1

− ˆy

)

∑

i=1

− ¯y

)

Here index i represents the generic data point of the

test-set, while y

and ˆy

denote the measured and fore-

cast values, respectively. Both RMSE and MAE mea-

sure the accuracy of the forecasting model on average:

RMSE is the squared root of squared errors mean,

while MAE averages absolute values of the errors. R

represents the proportion of variance in the dependent

variable that is predictable from the independent vari-

ables. Values range from 0 to 1, with 1 indicating per-

fect prediction and 0 suggesting no predictive power.

Each metric has its strengths: RMSE is sensitive to

large errors, MAE is more robust to outliers, and R

indicates overall model ﬁt.

In addition to the traditional KPIs, the value of the

forecasting methods has been evaluated by the new

measure, VPF. Let K denote the set of forecasting

techniques under evaluation. Each method k ∈ K

generates a supply proﬁle for each prosumer within

the REC, which then serves as input data in the opti-

mization model. By solving the optimization problem

|K | times− once for each forecasting method and as-

sociated production values G

— we evaluate the in-

ﬂuence of each forecast on decision outcomes. Let

represent the vector of decision variables when the

optimization model is run with forecast k. We then

calculate the objective function value when the so-

lution x

is applied, taking into account the actual

supply patterns. Forecast errors may require adjust-

ments to the initial operational plan, which can lead

to higher costs due to necessary interactions with the

balancing market. If for some time periods the real

PV generation is lower than the forecast one, the

aggregator may be required to purchase electricity

from the market at higher prices. On the contrary,

if the forecast overestimates the real proﬁle, the extra

amount should be fed back to the grid. Therefore, im-

plementing the solution x

under real conditions may

result in added imbalance costs. We denote the total

cost after these adjustments by z(x

To benchmark the results, we consider a scenario

of perfect forecast, which yields an objective function

value denoted by z

∗

. Thus, for each k, the VPF is

The Value of Perfect Forecasting in Optimizing the Management of Energy Communities

181

deﬁned as:

V PF

= z(x

) − z

∗

This metric enables a direct comparison between each

forecast approach and the perfect forecast scenario,

highlighting the ”cost” of forecast inaccuracy in terms

of suboptimal decision-making within the REC. A

lower VPF indicates that the forecast method pro-

duces results closer to the ideal solution under per-

fect information, suggesting a higher practical util-

ity in the decision-making process. The experiments

presented in Section 6 show how the new index may

complement the standard KPIs.

5 TEST CASE AND DATA

SETTING

To evaluate the impact of forecasting on the oper-

ation of RECs, we have considered a simple, yet

meaningful, case study related to a small community

composed of three types of members. Two of them

are prosumers, equipped with PV panels and BESs,

whereas the other is a simple consumer. Table 1 re-

ports the energy assets of the REC members, in terms

of number of PV panels and BES capacity, along with

the maximum operationing limits.

Table 1: Users’ energy assets.

Number Storage Grid Operation

Member panels Capacity limit

(kWh) (kW)

1 16 6 4.5

2 32 12 6

3 - - 3

Demand and tariffs have been assumed to be known.

The former has been derived using the data from (AR-

ERA, 2024) and (Giordano et al., 2020), whereas the

electricity tariffs have been derived using the Ital-

ian Single National Price as reference (Gestore Mer-

cati Energetici, 2023). The model has been coded in

Python and solved by the commercial solver (Gurobi

Optimization, LLC, 2024).

As for the forecasting, experiments have been car-

ried out by considering a three-year data set, includ-

ing 23832 observations, with a resolution of 1-hour.

The data have been provided by the University of

Calabria, Italy, and the following features have been

considered: Hour of the day (ranging from 1 to 24),

Relative Humidity (%), Temperature (°C), and Wind

Speed (km/h). The target value is the generation of

a module installed on the rooftop of one of the uni-

versity buildings

. Before running the tests, a clean-

The PV module, employing polycrystalline silicon

ing phase has been performed. Data points, for which

not all the features were available, have been removed

(735 data points). Night observations have been also

excluded from the dataset for preventing bias. At the

end of the cleaning phase, the ﬁnal number of data

records was reduced to 13462.

To improve the performance of the ML models,

a feature engineering process has been implemented.

This step focuses on deriving new, informative input

variables by transforming and processing the original

ones. This step starts by examining the correlation be-

tween the input variables and the target variable using

a correlation map, shown in Figure 1. This visualiza-

tion helps to identify the most relevant features and

potential redundancies or collinearities in the dataset.

Figure 1: Correlation matrix of the original features.

Analyzing the data, it appears that the feature

HourDay is one of the less signiﬁcant variables for

predicting power generation. However, a closer look

reveals that the relationship between power genera-

tion and the hour of the day is similar to a quadratic

function. To account for this, the squared values of

the hour of the day have been introduced as an addi-

tional input feature, improving the model’s ability to

capture this non-linear relationship.

Furthermore, as highlighted in the results of

(Nicoletti and Bevilacqua, 2024), the relationship be-

tween power generation and the meteorological vari-

ables— temperature and relative humidity— can be

more accurately modeled by incorporating temporal

variations. Speciﬁcally, the current values of these

variables, as well as their values one hour before and

one hour after, are included as input features. These

additional variables allow the model to better capture

the predictive power of short-term ﬂuctuations, which

are crucial for accurate forecasting. Finally, the most

technology is characterized by latitude: 39°21’N and Lon-

gitude: 16°13’E. The module’s surface presents an area of

1663 mm x 998 mm, and a tilt angle of 30°. It faces south-

east, with a nominal power of 245 W.

ICORES 2025 - 14th International Conference on Operations Research and Enterprise Systems

182

Figure 2: Power production for an autumn day: predictions

versus perfect forecast.

recent historical values of power generation are in-

cluded as a critical input feature. At the time of fore-

casting, the most recent available data for each hour

comes from 48 hours prior. Incorporating this histor-

ical information ensures that the model exploits tem-

poral continuity and accounts for persistent trends or

patterns in power generation.

Data processing, including feature engineering

and the application of ML has been implemented in

Python. Data have been split in the training and the

test sets with a ratio of 15%.

6 NUMERICAL RESULTS

This section is devoted to the presentation and discus-

sion of the numerical results. Figure 2 shows the fore-

cast power production obtained considering the dif-

ferent ML techniques for a given day in November. In

the same Figure, the perfect forecast, representing the

actual power production, is also included. As shown,

all the techniques provide predictions that capture

the general trend of the real production values; how-

ever, differences in accuracy and behavior are appar-

ent among the methods. Some techniques closely fol-

low the actual production, especially during peak pro-

duction periods, while others exhibit more signiﬁcant,

even though limited, deviations. Quantitative perfor-

mance metrics further highlight the accuracy of each

technique. As summarized in Table 2, RF method

consistently outperforms the other methods, achiev-

ing the lowest values for RMSE and MAE, as well

as the highest of R

. The STK method demonstrates

comparable performance to the RF, suggesting that it

effectively combines the strengths of multiple models,

although it does not signiﬁcantly outperform the best

individual technique (RF in this case). On the other

hand, the GBRT method performs the worst among

the tested techniques underlining the limitations in its

ability to capture the underlying patterns.

Forecast data have been used as input for the op-

timization problem presented in Section 3. The prob-

lem is solved iteratively, using each time the forecast

Table 2: Accuracy evaluation via traditional KPIs.

Methods/KPIs RMSE MAE R

SVR 23.16 13.20 0.89

RF 20.06 12.15 0.91

ADA 22.76 15.70 0.89

GBRT 23.44 14.95 0.88

STK 21.91 13.29 0.90

obtained with a different technique. The resulting

daily strategies, represented in terms of energy trans-

action with the main grid, have been used to deter-

mine the VPF. Imbalances, evaluated with the respect

to the strategy suggested when considering the per-

fect forecast, are corrected by recurring to the balanc-

ing market. To illustrate this process, Figure 3 depicts

how imbalances are managed when forecasts from the

RF technique are used. The ﬁgure highlights the ini-

tial transaction strategy based on RF predictions and

the subsequent adjustments required when real values

are observed.

Figure 3: REC transactions with the grid.

To gain further insights, the analysis was extended

by running the model for representative days across

different seasons. This seasonal analysis revealed the

impact of weather and environmental variability on

forecasting accuracy. Table 3 reports the values of the

VPF measure (in percentage) for the different tech-

niques and the different days.

Table 3: VPF (%) for the different ML techniques for dif-

ferent days.

Winter Spring Summer Fall Mean

SVR 0.52 1.22 4.74 7.97 3.61

RF 0.16 1.12 2.43 4.78 2.12

ADA 0.00 0.98 2.83 2.25 1.51

GBRT 0.76 1.09 3.30 7.60 3.19

STK 0.27 1.08 3.49 4.51 2.34

Looking at the results some considerations can

be drawn. First of all, we may observe that for all

the techniques, ”Fall” presents the higher VPF val-

ues, indicating higher variability potentially related to

instability in weather conditions. This may be due

to unpredictable weather patterns, such as rapid tem-

perature changes or varying cloud cover, which in-

The Value of Perfect Forecasting in Optimizing the Management of Energy Communities

183

crease forecasting difﬁculty. On the contrary, ”Win-

ter” shows the best values reﬂecting very likely more

stable conditions easier to predict. When comparing

the techniques for a given reference day, we may note

that ADA provides the best results followed by the

RF and the STK. Thus, while ADA is not competitive

when evaluated via traditional KPIs, when consider-

ing the VPF seems to provide the best results.

The results presented further emphasize an impor-

tant observation: the forecasting technique that per-

forms best under conventional metrics may not al-

ways be the most effective for prescriptive purposes,

especially in contexts where accurate adjustments and

decision-making are critical. This distinction high-

lights the need to tailor performance evaluations to

the speciﬁc application or domain requirements.

7 CONCLUSIONS

This paper focuses on the evaluation of forecasting

techniques from a prescriptive perspective. Speciﬁ-

cally, the study applies ﬁve ML techniques to train

predictive models for PV generation forecasting,

which are then used as input parameters for an opti-

mization problem aimed at deﬁning the optimal daily

operational strategy for a REC. The different tech-

niques are evaluated by using both standard accuracy

metrics and a new measure, the VPF, used to mea-

sure the cost incurred to adjust operational plan for

compensate for the deviations between forecast and

actual values. The preliminary results seem to point

out that the ML model with the highest score on the

standard statistical metrics is not necessarily the most

effective for optimization purposes. This study under-

scores the need for a more holistic approach to evalu-

ating forecasting models, especially when they are in-

tegrated into optimization workﬂows. By considering

both standard metrics and application-speciﬁc indices

like VPF, stakeholders can select models that are not

only accurate, but also cost-effective for operational

decision-making.

In this sense, the present work constitutes a pre-

liminary investigation within this ﬁeld of application.

Firstly, it is recommended that the insights provided

by the new index introduced in this study should be

further replicated through a range of different com-

putational experiments and application settings. This

will demonstrate the practical utility and generalis-

ability of the proposed index in different contexts.

Future research directions could focus on achieving

a deeper integration between predictive and prescrip-

tive processes. One promising approach may be em-

bedding the prediction process within optimization

models by leveraging innovative approaches like con-

straint learning. In alternative the training process of

predictive models may be carried out, taking in ac-

count the structure of the optimization problem, like

in ’Smart Predict, Then Optimize’ framework.

ACKNOWLEDGMENTS

We acknowledge the ﬁnancial support from: PNRR

MUR project PE0000013-FAIR.

REFERENCES

ARERA (2024). Analisi dei consumi domestici, pre-

lievo medio orario per provincia, 2021- cosenza.

https://www.arera.it/dati-e-statistiche/dettaglio/

analisi-dei-consumi-dei-clienti-domestici [Accessed:

(31/10/2024)].

Breiman, L. (2001). Random forests. Machine Learning,

45.

Breiman, L., Friedman, J., Stone, C., and Olshen, R. (1984).

Classiﬁcation and Regression Trees. Taylor & Francis.

Ding, X., Liu, J., Yang, F., and Cao, J. (2021). Ran-

dom radial basis function kernel-based support vec-

tor machine. Journal of the Franklin Institute,

358(18):10121–10140.

Drucker, H. (1997). Improving regressors using boosting

techniques. In Icml, volume 97, page e115. Citeseer.

Friedman, J. (2001). Greedy function approximation:

A gradient boosting machine. Annals of Statistics,

29:1189–1232.

Gestore Mercati Energetici (2023). National hourly elec-

tricity selling prices. https://gme.mercatoelettrico.org/

it-it/Home/Esiti/Gas/IGIndex/Statistiche/Sintesi [Ac-

cessed: (19/10/2024)].

Giordano, A., Mastroianni, C., Scarcello, L., and Spezzano,

G. (2020). An optimization model for efﬁcient energy

exchange in energy communities. In 2020 Fifth Inter-

national Conference on Fog and Mobile Edge Com-

puting (FMEC), pages 319–324.

Gurobi Optimization, LLC (2024). Gurobi Optimizer Ref-

erence Manual.

Kodaira, D., Tsukazaki, K., Kure, T., and Kondoh, J.

(2021). Improving forecast reliability for geograph-

ically distributed photovoltaic generations. Energies,

14(21).

Ma, M., He, B., Shen, R., Wang, Y., and Wang, N. (2022).

An adaptive interval power forecasting method for

photovoltaic plant and its optimization. Sustainable

Energy Technologies and Assessments, 52:102360.

Nicoletti, F. and Bevilacqua, P. (2024). Hourly photovoltaic

production prediction using numerical weather data

and neural networks for solar energy decision support.

Energies, 17(2).

ICORES 2025 - 14th International Conference on Operations Research and Enterprise Systems

184

Peterssen, F., Schlemminger, M., Lohr, C., Niepelt, R.,

Hanke-Rauschenbach, R., and Brendel, R. (2024). Im-

pact of forecasting on energy system optimization.

Advances in Applied Energy, 15:100181.

Putz, D., Gumhalter, M., and Auer, H. (2023). The true

value of a forecast: Assessing the impact of accu-

racy on local energy communities. Sustainable En-

ergy, Grids and Networks, 33:100983.

Ruszczy

nski, A. and Shapiro, A. (2003). Stochastic Pro-

gramming, Handbook in Operations Research and

Management Science. Elsevier Science, Amsterdam.

Sharadga, H., Hajimirza, S., and Balog, R. S. (2020).

Time series forecasting of solar power generation for

large-scale photovoltaic plants. Renewable Energy,

150:797–807.

Vapnik, V. (1998). The Support Vector Method of Function

Estimation, pages 55–85. Springer US, Boston, MA.

Wolpert, D. H. (1992). Stacked generalization. Neural Net-

works, 5(2):241–259.

Yao, S., Pan, L., Yu, Z., Kang, Q., and Zhou, M. (2019). Hi-

erarchically non-continuous regression prediction for

short-term photovoltaic power output. In Proceedings

of the 2019 IEEE 16th International Conference on

Networking, Sensing and Control, ICNSC 2019, pages

379–384.

The Value of Perfect Forecasting in Optimizing the Management of Energy Communities

185