3.1. Goal and Scope
In
Table 2, studies are classified for the goal of using a DLCI in three kinds of alternative assessments: retrospective assessments, scenario analyses, and forecasts. Depending on the chosen temporal scope, retrospective assessments can include periods of time that are in the future compared to the time of the assessment. Retrospective studies assess past or current systems as they are, while scenario analyses typically rely on what-if simulations, which diverge from the past course of events. Furthermore, unlike scenario analyses, forecasts aim to predict elements of the assessed future.
For each of the three types of assessments, the studies are divided between two alternative categories regarding the data collection. The first category uses previously collected DLCI data. They are not necessarily related to the past but they are already fully available at the time of the assessment. In the second category, DLCI data are continuously collected and so they are not fully available at the beginning of the assessment. Since the focus is on case studies rather than meta-studies, the four literature reviews are not included in this classification.
Firstly, the majority of studies, 37 out of 63 publications, deal with retrospective assessments. These are either methodological publications [
21,
22,
23] or focus on determining the dynamic environmental impacts of a system as it is [
16,
24,
25]. The development of DLCI methodologies is crucial, and so is the dynamic assessment of systems which could lead to a bottom-up approach for a fully dynamic LCI database, rather than the top-down approach used by Pigne et al. [
26]. These publications might not be at the center of attention for researchers aiming to develop DLCI models for Industry 4.0 applications, yet we consider them in the review to provide a fuller description of the literature.
Of these 37 publications, 10 deal with continuously collected DLCI data. The manufacturing field is represented by various publications regarding personalized furniture manufacturing [
27], a generic shop floor [
4,
28], injection molding [
29], grinding [
30], and ceramic tile production [
31]. The relative overrepresentation of manufacturing case studies in the continuously collected DLCI data category can be attributed to the focus that this sector has received in Industry 4.0 initiatives [
6]. Consequently, we can assume that sensors used for relevant data collection are more widely adopted than in other sectors. Nonetheless, case studies in this category include power plants [
32], a generic enterprise case study [
20], and solar photovoltaic systems [
33]. Of these 10, eight are conference proceedings and only two are journal articles [
30,
31]. Therefore, some of the publications keep the description of the models at a high level [
20,
29,
33]. Others, even if they are more detailed, omit key information such as choices regarding allocation [
4,
31]. The only submitted supplementary data allow for the reproduction of results, but no code is available to reproduce the overall experiment for other case studies [
30].
Table 2.
The relationship between the type of the assessment (rows) and the data collection period in time (columns).
Table 2.
The relationship between the type of the assessment (rows) and the data collection period in time (columns).
Type of Assessment | Previously Collected DLCI Data | Continuously Collected DLCI Data |
---|
Retrospective | [16,17,18,19,21,22,23,24,25,26,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50] | [4,20,27,28,29,30,31,32,33,51] |
Scenario analysis | [52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69] | [70,71,72] |
Forecast | | [73,74,75,76,77] |
Secondly, 20 publications have as a goal the assessment of scenario analyses. Here, hypotheses are tested by changing key parameters in the model. Examples span from electric vehicles charging scenarios [
52], cost-optimized lithium-ion battery management [
56], and configuration changes in a manufacturing process [
59], to future energy system optimization [
62,
65]. A particular case of scenario analysis limits the change among scenarios to the temporal differentiation parameter, leaving the product system parameters constant [
67].
Only three publications use continuously collected DLCI data to perform scenario analysis [
70,
71,
72]. Curiously, one of them is the oldest publication among those describing continuous data collection [
70]. Here, MTConnect enables continuous data collection from the shop floor, which is then used as input to discrete event simulation models [
70]. The aim is to project DLCI data and consequently environmental impacts, similarly to what Rodger et al. [
64] envision without the continuous data collection. The second publication assesses four environmental indicators for the hourly electricity consumption in France [
71]. These are then used to assess the environmental performance of demand-side management programs, finding that including a carbon price in the assessment might improve the performance [
71]. The third publication of this group allows for both retrospective assessments for certification purposes and scenario analyses [
72]. The user can therefore visualize LCA results across months of data, analyzing contributions toward the impact of a product within a timestep and over time as well as the impact of a selected driver across impact categories [
72]. In the scenario analysis, the user can test the effectiveness of an action towards impact reduction goals over time and across impact categories [
72].
Lastly, four papers use continuously collected DLCI data to make forecasts. They adopt machine learning to forecast the electricity technology mix [
73] or the environmental impact of electricity consumption [
74,
76,
77] of the following day. Even if these papers mainly aim to demonstrate the potential benefits of future applications using previously collected DLCI data, such applications would require continuously collected data. Indeed, the forecast window would shift and the machine learning models would be retrained to include new data on a daily basis.
A different classification of the goals found in the literature is detailed in
Table 3. A more precise description of each goal is provided during the description and discussion of
Table 3, later in this section. Even though some goals overlap with the findings of
Table 2, they do so only partially. For example, goal 5 only partially overlaps with the continuously collected DLCI data category that we previously discussed. For each goal, we differentiate publications based on the chosen framework: attributional, consequential, or both attributional and consequential. Attributional LCA assesses the potential environmental impacts of a product or service system [
78]. Consequential LCA assesses the potential environmental impacts due to a small change in demand for the studied products and system boundaries only include those activities that react to such change [
79]. The number of papers referenced is not 63 as this classification is not exclusive, so publications can have more than one goal. Literature reviews are not included in this assessment, as they would cover most goals. Therefore, no publications were found to have more than two goals.
The more publications deal with a goal, the higher the goal is in the list. The difference in publication number between the first and the third goal is only four, depicting a fair distribution of goals within the literature. Goals 4 and 5 are less common, with only 13 and 12 publications, respectively, and even less so is goal 6, which is limited to three publications.
Goal 1 is the most generic of the classification, as it groups publications that are not aligned with the other, more specific goals. Due to the inclusion and exclusion criteria used in this systematic literature review, all publications use DLCI. Moreover, all publications in this goal use attributional modeling. Some studies aim at assessing the dynamic environmental impact of a product system such as integrated urban water systems [
38], electric vehicles [
18,
53], residential buildings [
54], power-to-X technologies [
61], or solar photovoltaics [
63]. Other publications put less focus on the case study and on the dynamic assessment results and more on the methodological approach that is developed. Examples span from a sensitivity analysis on the temporal resolution [
55] to agent-based modeling of a biomass supply chain [
41] to a computational tool for DLCI [
23]. Further discussion on these methodological publications can be found in
Section 3.2. Both attributional and consequential approaches are used with regard to goal 1 for the modeling of different electricity systems on an hourly basis, which are then applied to electric car charging in Portugal [
39] and in the USA [
16] as well as to space heating in buildings in France [
48].
To be included in this goal 2 class, publications need to clearly describe how insights gained with DLCI could be used to pursue environmental impact reductions. Indeed, some publications state among their aims the definition of a sustainability-based optimization, but when the optimization is left as a proof of concept these works are excluded from goal 2 [
4,
27]. Within goal 2, nine publications use attributional modeling and seven use the consequential one, the most of any goal class. This is understandable, given the need to understand potential unforeseen consequences of decisions aimed at reducing environmental impacts. Still, some publications propose strategies, based on attributional modeling, to reduce environmental impacts based on hotspot identification [
30], fuel purchasing and operational strategies [
32], discrete-event simulation of manufacturing systems [
59,
64], decision support for material and process innovation [
40], as well as shifting production towards hours that are forecast to have a less impacting electricity mix [
73,
74,
76]. Considering consequential LCA publications, a trade-off between different optimization objectives, such as water consumption and the impact on climate change, is quantified in the case of an industrial wastewater treatment system [
24]. Moreover, shifting electricity demand based on forecast marginal hourly electricity mixes is proposed for energy storage systems [
56], cloud computing [
43], data centers [
49], demand-side management [
69], and generic electricity consumption [
77].
Two thirds of the publications in the goal 2 class deal with a single optimization objective: reduction in energy consumption [
75] or reduction in climate change impacts [
32,
43,
49,
56,
64,
73,
74,
76,
77]. By doing so, these publications ignore the risk of potential burden shifting, which happens when impacts are reduced in one impact category but increased in one or more others. Multi-objective optimizations deal with relevant flows and midpoint impact indicators such as monthly electricity use, electricity use per produced bearing unit, monthly climate change impacts, and climate change impacts per produced bearing unit in the case of Lofgren et al. [
59]; climate change impacts and water consumption in O’Connor et al. [
24]; and 11 midpoint impact categories from the Environmental Design of Industrial Products 97 (EDIP 97) [
80] in Filleti et al. [
30]. In the case of net-zero energy buildings, a multi-objective optimization focuses on two indicators: construction costs and impact on climate change [
57]. The temporal demands for electricity and heating are optimized to reduce impacts on climate change. While the difference between optimums obtained with a static LCA and a dynamic attributional one is found to be minor, they become more relevant when dynamic consequential and attributional optimums are compared [
57]. Finally, in the study by Walzberg et al. [
69] the multi-objective optimization deals with four endpoint impact categories. However, the use of endpoint impact categories is not recommended for comparison tasks, implicit in optimization studies, because trade-offs and burden shifting within an endpoint impact category cannot be tracked [
81].
A common objective in the literature is the comparison of the environmental impacts of a product system using static and dynamic LCIs, classified under goal 3. This is often used to justify the added complexity of a dynamic study against the traditional static LCA [
46]. The need for such a demonstration is exacerbated by the fact that ISO standards do not cover dynamic assessments [
1,
2], which might increase the perceived arbitrariness of dynamic LCA results. Olindo et al. [
45] argue for the need for DLCI to assess systems such as battery electric vehicles and electrolytic hydrogen, while highlighting the limitations of guarantee of origin and residual electricity mix [
45]. DyPLCA is a tool to extend the temporal differentiation to the whole LCI, based on ecoinvent 3.2 [
26]. This top-down approach can help to solve issues with DLCI, which is typically temporally differentiated in the sole foreground system. Based on the same original model, [
23], Shimako et al. [
67] perform a sensitivity analysis on parameters such as the temporal resolution of different impact categories [
67]. Electricity systems are again at the center of the case studies in two publications that employ both modelings under goal 3 [
21,
36]. Here, the discussion verges on the how and why each modeling approach should be used, while at the same time dealing with the choice between a static and a dynamic study. In the Italian case, hourly marginal electricity mixes obtained with the consequential approach vary more than average hourly electricity mixes [
21]. However, hourly marginal electricity mixes are suggested for applications regarding buildings with variable electricity demand [
36] and demand-shifting applications in general [
21]. Other methodological articles that are included in this publication will be further discussed in
Section 3.2.
Publications classified under goal 4 deal with the assessment of the environmental impacts of future scenarios, relative to the time of the assessment. Depending on the goal and scope of the study, the assessment can continue for hours, days, or decades. It is important to mention that this is not strictly correlated to the time horizon that is chosen in the LCIA phase, which is less variable. Rather, it is related to the temporal scope, which is usually linked to the life cycle of the product [
5]. However, the temporal scope is usually limited to the short-term for discrete event modeling of manufacturing systems [
64,
70], but it can span to the medium-term [
72] and long-term for assessments that consider larger and slower changes. For example, in the case of the assessment of the penetration of electric vehicles in the Italian fleet, the temporal scope lasts until 2030 [
65]. Therefore, the temporal scope should be carefully chosen depending on the the goal of the study and to minimize cut-offs due to this choice. Moreover, it is common for continuously collected DLCI data studies to see the temporal scope of the study shift with time [
73,
74,
76].
Publications classified under goal 5 investigate the potential of IoT and sensorization to automate the DLCI data collection. Due to the extensive time and costs required for the traditional LCI phase, being able to automate the most relevant sections could bring significant benefits [
4]. As noted in the previous classification regarding continuous DLCI data collection, manufacturing case studies are well represented [
4,
20,
27,
28,
29,
30,
31,
51,
70,
72]. It is worth mentioning how Barni et al. [
27] suggest the development of a digital twin based on real-time data collection, which can support decision making and in particular drive sustainability-aware decisions [
27]. Majanne et al. [
32] apply DLCI modeling to the monitoring of the environmental impacts of monitored power plants [
32]. A deeper dive into the methodologies of the publications under this goal can be found in
Section 3.2.
The comparison of the dynamic environmental impacts of different product systems, which fulfill similar functions, is present in three articles only, as classified in class 6. Di Florio et al. [
37] compare two cogeneration systems linked to thermal energy storages [
37]. While daily and seasonal variability of the DLCI is presented and discussed, the comparison of environmental impacts between the systems is between seasonal averages. To test the robustness of the results, four Monte Carlo simulations, one for each season, are carried out [
37]. Faria et al. [
17] perform a comparative DLCA among gasoline, diesel, hybrid electric vehicles, and battery electric vehicles [
17]. Comparisons of environmental impacts are made using yearly and monthly averages, without discussing the distribution of the environmental impacts within each year or month. Finally, Shahraeeni et al. [
66] compare diesel and natural gas for light duty trucks using static results [
66].
About 78% of the case study publications, 49 out of 63, use attributional LCA, seven choose the consequential approach, and seven publications adopt both of them. Notably, no publication uses consequential modeling under goals 5 and 6. This is understandable for goal 5, as in these publications the automation of the data collection is described but the focus is not on the decisions that new available data could enable. However, a consequential assessment of the use of one technology rather than another, as is investigated under goal 6 publications, would seem preferable to support a more robust choice among the alternatives.
3.2. Integration of Dynamic Life Cycle Inventory
In this section, we analyze the different strategies that are found in the literature to integrate a DLCI into LCA studies. To conduct this analysis, we use three of the data extraction forms described in
Section 2.4, namely the modeling, the dynamic component in the LCI, and the temporal resolution. When considering the three data extraction forms concurrently, we are able to shed light on how previous studies have dealt with specific problems. Finally, we describe the methodologies that, for different reasons, could be relevant for practitioners working with DLCI.
The modeling is divided in three exclusive categories, namely white-box, gray-box, and black-box [
82]. White-box models, such as most LCA studies, are fully explainable and interpretable, meaning that results are obtained through known functions of the inputs [
82]. Therefore, the dynamic flows are synthesized and modeled into fixed equations with the aim of representing their temporal variability. As with every white-box model, the abstraction due to the use of fixed equations might lead to lack of precision in the results, especially when the model is not robustly validated. Black-box models such as machine learning are not explainable and interpretable, so it is not possible to reproduce step-by-step how the inputs are used to result in the outputs of the model [
82]. Machine learning is increasingly popular for a multitude of tasks, due to the growing availability of data and computational power, and is touted as being able to support environmental issues in various ways [
83]. Gray-box models are a combination of the theoretical structure of white-box models and of the data-driven approach of black-box models.
The dynamic component of the LCI, the second data extraction form used in this section, is relevant as it provides a more detailed picture than the overall case study. This focus helps to identify which processes or sub-systems vary with time, as most publications do not have a fully dynamic LCI. The third data extraction form is the temporal resolution used in the DLCI. Of particular interest is its link to the dynamic component for the LCI, as it can be a suggestion to practitioners on how to deal with a specific problem.
In
Table 4, the modeling is the means by which the dynamic feature is integrated in the study. In 57 publications, or about 93% of the reviewed literature, after the exclusion of the four literature reviews, a white-box model is chosen. Cornago et al. [
73] is the only study to deploy a gray-box modeling approach [
73]. The objective function of the machine learning models, the black-box component, is the electricity production of each technology present in the local mix. The machine learning models are deep neural networks, which use as input the electricity production of each technology from the previous day. The forecasts are then linearly combined with their respective LCA impact in the white-box component, to obtain the hourly LCA impact of the electricity mix [
73]. Black-box models found in the literature deal with the same forecasting problem, but choose a strategy called direct forecasting [
73], as it does not require a subsequent linear combination. Indeed, the objective of the machine learning models in black-box studies is the hourly LCA impact of the electricity mix [
74,
76,
77], as well as electricity consumption [
74,
76]. The inputs of the forecasting models are similar in nature to those by Cornago et al. [
73]. However, machine learning models vary from a neural network that counts two layers of long short-term memory (LSTM) cells followed by a layer composed of a linear output neuron [
74,
76] to a support vector machine [
77].
Electricity consumption, electricity technology mix, and the combination of the two represent the largest share for the dynamic component in the LCI within the selected literature, contributing 4, 13, and 22 studies, respectively. This can be traced to the relevance that electricity consumption has as a driver of LCA impacts, as well as the maturity of electricity meters and the variability of the environmental impacts of electricity mixes. Considering the current trend of electrification of various sectors of the economy, most notably in the automotive sector, the attention seems like a good bet.
However, a more holistic approach that extends the dynamic components in the LCI as much as possible should be preferred, and it should be made possible with a wider data collection. A choice adopted by 13 publications is to consider as the dynamic component all flows that are part of the foreground system, which are those that can be measured and controlled. This solution is adopted predominantly by manufacturing case studies, where a gate-to-gate perspective is common [
30,
70].
Even more comprehensive is to use a fully dynamic LCI, where the whole product life cycle system is dynamic. Three methodologies, all white-box models, are developed to manage the computation of fully dynamic LCI [
22,
23,
35]. They are designed in a way that can be integrated with a dynamic LCIA.
Firstly, the enhanced structure path assessment (ESPA), which extends structural path analysis, is a widely known technique in input–output analysis [
35]. It makes use of power series expansion to solve the dynamic inventory, and the matrix inversion is replaced with a product of convolution of the discrete distribution functions. A drawback of this approach is that it is not available online and is used in just one other publication [
46].
Secondly, the approach that consists of a direct traversal of the supply chain graph [
23] introduces a promising method for dynamic LCI that has been developed as a prototype web application, called DyPLCA [
26]. It is based on a process flow network structure and makes use of a depth-first graph search algorithm to build the temporal model. However, this proof of concept is not coupled with an LCIA framework and it is not clear if the method can deal with datasets without temporal information, raising doubts over its integration potential with existing LCA databases. Regarding the treatment of the LCI as a graph, it is worth mentioning that this approach poses a key methodological challenge due to the cyclic nature of the supply chain graphs. Loops can be encountered, and a cut-off function must be applied to halt potentially infinite loops in supply chain traversal. This method is used in two other publications to assess the relevance of using DLCI and the assumption on the time horizon on environmental impacts results [
67] and to describe the online tool that enables the use of the methodology [
26], which is available to the public.
Finally, Cardellini et al. [
22] employ a best-first search strategy to solve the dynamic inventory problem, creating the Temporalis tool [
22]. Here, processes are nodes of the graph and flows are edges, which can be both static and dynamic. Furthermore, the temporal distributions of the product–process and biosphere–process interactions are taken into account, using convolution as used by [
35]. This methodology therefore builds on top of the first two, while at the same time solving some shortcomings. Namely, this approach can deal with LCI and LCIA that are only partly dynamic, as is often the case, and can manage both relative and absolute time references. The temporal resolution can be as little as 1 s. A major advantage is that the Python code is fully available in a dedicated GitHub folder.
The ESPA, DyPLCA, and Temporalis are designed for DLCA using previously collected data. Other models, that we are about to describe, might be better suited to elaborate continuously collected data, since they were developed with this scope. We should stress, however, that, if a dynamic LCIA is needed, the three previously analyzed models and tools should be the starting point for future developments.
Before the development of models for continuously collected DLCI data is possible, practitioners should consider in which format data are generated from the different machines. A solution to make such formats uniform is the use of a common standardized interface protocol for data exchange and collection from the shop floor, such as MTConnect [
4,
30,
51,
70]. Subsequently, attention should be devoted to data management steps to ensure efficient data handling, storage, and analysis [
4].
If the model needs to deliver real-time emission and environmental impacts results, continuous DLCI data collection might not be enough. Indeed, needed data might not be available [
4] or measurements might be unreliable without a certain time lag [
32]. Simulation methods should be employed to overcome these issues. For example, Majanne et al. [
32] apply extended Kalman filters to estimate real-time fuel consumption used in a power plant [
32].
Bengtsson et al. [
70] develop a discrete-event simulation of a manufacturing system whose input is data gathered with MTConnect [
70]. The goal is to project long-term environmental impacts based on production statistics. Therefore, the simulation conserves the modeling granularity obtained by the MTConnect data, both in terms of assessed dynamic processes and in terms of temporal resolution. The use of discrete-event simulation for DLCI is also advanced by Lofgren et al. [
59] and, more recently, Rodger et al. [
64], although these publications do not rely on continuously collected DLCI data [
59,
64]. To represent the manufacturing system, a unidirectional process flow chain, alternated by buffers, is proposed [
64]. Buffers are used to model time lags, material availability, and process bottlenecks [
64]. A similar approach is a modular DLCI model for real-time assessment of the dynamic environmental impacts of a manufacturing system [
28]. A gate-to-gate approach is chosen and the granularity of the modules is a trade-off between the ease of modeling and the detail of the assessment. Potential issues regarding the allocation of environmental impacts are discussed and recommendations for how to deal with them are presented for cases such as waste, discarded products, standby times, setup and ramp-up times, transports, storage, energy systems, heating, ventilation, air conditioning, buildings, administration, and auxiliary systems [
28].
To widen the consideration of DLCI beyond a manufacturing floor gate-to-gate approach, an interesting solution is proposed by Tao et al. [
20] and Ferrari et al. [
31]. Here, continuous DLCI data collection combines IoT infrastructure with enterprise resource planning (ERP) software. Thus, the ERP provides dynamic data for sections of the product system that would have otherwise been part of the static background system. However, the resulting data quality might not be as high as it is for primary data, due to difficulties in characterizing processes mapped through the ERP [
31]. Rovelli et al. [
72] do not have access to ERP data in their steel making case study and have to collect primary data from various sources [
72]. The modular approach deals with collection and wrangling of data into an Excel spreadsheet, which is subsequently manipulated in Python to allow the integration of different functions such as label-specific sustainability reporting, contribution analysis, what-if scenarios, and data visualization tools [
72].
Collet et al. [
55] propose a methodology to assess which flows, part of the product system, are more relevant when introducing a partially dynamic LCI [
55]. Through sensitivity analyses on environmental and economic flows, the dynamic parts of the LCI should be those that, with the added dynamic, translate to higher environmental impact variations. Moreover, a different temporal resolution is suggested depending on which impact category is the most affected [
55]. The goal of this article seems promising for the smart use of a partially dynamic LCI. However, the methodology requires significant data collection or assumptions and it is not adopted by other publications.
Electricity-related studies either use one hour or half an hour as temporal resolution, due to the prevalence of this temporal resolution in electricity network operators’ databases. Indeed, depending on the country, these temporal resolutions are used in day-ahead electricity markets. Day-ahead markets determine how much electricity should be produced in a certain hour by the power plants available in the market zone, as well as imports and exports with neighboring market zones. Therefore, they establish the hourly electricity technology mix trajectory for the following day.
When selecting the temporal resolution for a study with dynamic LCIA, practitioners should consider which impact category could be impacted the most. Indeed, Shimako et al. [
67] demonstrate with a series of sensitivity analyses on the temporal resolution, one for each impact category, that different LCIA models have different sensitivities to the temporal resolution parameter [
67]. While the climate change impact category does not seem to be influenced by temporal resolutions up to 1 year, the toxicity model shows a greater response, thus requiring a temporal resolution of 0.5 day [
67].
3.3. Identification of Research Gaps
In this section, we summarize and discuss the research gaps highlighted by this systematic literature review, as listed in
Table 5. Therefore, this section aims at guiding future research development for DLCI projects. For each research gap, we describe its meaning and we discuss its relevance. However, the order presented in
Table 5 does not reflect a ranking of significance, but rather it follows the flow of the presentation of
Section 3.1 and
Section 3.2.
The first two research gaps are the most directly linked to sustainability-linked strategies brought forward by Industry 4.0. Indeed, only 17 publications deal with continuous DLCI data collection and only 11 automate the data collection process through IoT or sensors in general (see
Table 2 and
Table 3). This practice would have the potential to generate large amounts of data, which could be used to compile DLCI databases, to improve the representativeness of static LCI databases, and to design new strategies to decrease environmental impacts. More studies are needed to evaluate and compare proposed methods to define best practices. The literature is even more lacking when it comes to scenario analyses or forecast using continuously collected DLCI data. One of the pillars of Industry 4.0 is the CPS, which allows for the dynamic and autonomous coordination of machines and production systems, which share information [
6]. To be able to include life cycle engineering strategies in the CPS goal, sensitivity analyses and forecasts are both needed to support the decision making.
The assessment and management of uncertainty are well-debated and common in the general field of LCA, as it is considered crucial to improve data quality, reliability, and credibility of LCA results [
84]. However, the topic is rarely addressed in the DLCI literature, although an approach aiming to reduce uncertainties in the carbon footprint of solar photovoltaic systems is proposed [
33]. The only instance of a Monte Carlo simulation of a case study in the literature performs the analysis at a seasonal level, through four static simulations [
37]. While Monte Carlo simulations in DLCI might be computationally unfeasible, further research is needed to adapt known methodologies or to develop new ones.
The interpretation phase is often considered overlooked and some of its steps are not commonly applied by LCA practitioners [
85]. Moreover, the time dimension adds complexity to the interpretation phase, which results in different strategies to communicate results within the literature. Examples vary from a standard line plot [
22], to histograms with distribution ranges [
43] and to box plots [
19]. An important variable in this choice is the presence of dynamic LCIA characterization factors. Moreover, depending on the goal of the study, on the temporal resolution, and on the time horizon, using static cradle-to-gate LCA coefficients might lead to incorrect variability of impacts over time. An example is in the assessment of the environmental impacts of the consumption of electricity on an hourly basis. To account for the overall impact, static cradle-to-gate LCA coefficients can correctly characterize each technology in the electricity mix. When considering impact variability for consumption load shifting from one hour to the next, however, only the operational impacts should be contemplated. Further research is needed to provide customized guidance on the interpretation phase of DLCI studies, as Laurent et al. [
85] do for static LCA ones [
85].
Static LCA strategies to reduce impacts typically revolve around hotspot analysis [
86]. However, hotspot analysis can be adapted to DLCI studies to support the use of the increased dynamic data availability [
87]. Any strategy should take into account potential burden shifting across impact categories, life cycle stages, and, when DLCI is used, through time.
The literature is particularly concentrated on strategies that take advantage of the variability of electricity mixes and consumption. However, a systematic effort should be dedicated to investigate which sectors could pursue environmental impacts reduction by using temporally differentiated data.
The last research gap might be one of the most challenging to address. This could be due to privacy or business concerns in sharing the large amounts of data produced within DLCI studies. Indeed, partial non-disclosure agreements are common for LCA practitioners even in static LCA, as industrial partners prefer to publish aggregated results. However, methodologies should be shared with as much detail as possible, to ensure replicability and transparency of the scientific process, as well as to verify sustainability claims.