KezdőlapEnglishMunicipal solid waste prediction: 36 models and 455 variables – what the...

Municipal solid waste prediction: 36 models and 455 variables – what the University of Pannonia review reveals

If you like our site, mark us as a preferred source on Google — so you’ll see our articles more often in search!

★Mark us as a preferred source

How much waste will a city generate next year, and how much of it will end up at the sorting plant, the incinerator or the landfill? Without municipal solid waste prediction, neither the reorganisation of a collection district nor the capacity planning of a treatment plant can rest on solid ground.

Researchers from the Sustainability Solutions Research Lab at the University of Pannonia in Veszprém, Hungary – Róbert Fejes, Róbert Kurdi and Viktor Sebestyén – reviewed 78 scientific studies in a paper published in Resources, Conservation & Recycling Advances, and used them to build a decision-support framework that recommends a model structure based on the analysis goal and the available data.

In short: The most suitable model for municipal solid waste prediction depends primarily on how many observations and how many explanatory variables are available, and on whether an operational or a strategic decision needs to be supported. The researchers systematised 140 model applications from 78 studies: artificial neural networks are the most common, while economic and demographic indicators dominate the 455 variables identified.

Why waste volumes need to be anticipated

According to the World Bank’s estimate, global municipal solid waste generation could rise from 2.01 billion tonnes in 2016 to 3.40 billion tonnes by 2050. Changes in quantity and composition directly affect whether a given collection system or treatment technology remains fit for purpose. Good forecasting is therefore not an academic exercise: it is the basis for planning collection rounds, transport and treatment capacity, and for calculating operating costs realistically. The authors also see a role for forecasting in public awareness, as quantified projections make the environmental and economic consequences of waste generation visible.

The study starts from a framework developed by the authors, which they call the GCSIL model. The letters stand for five elements of the waste management chain: generation, collection, sorting, incineration and landfilling. The framework separates two decision processes:

  • Short-term operation model: supports the day-to-day running of collection and sorting. It is fed by dense, long time series from waste management companies – quantities per fraction, data per collection vehicle and technological parameters.
  • Long-term decision model: supports strategy, infrastructure and the choice of treatment routes such as incineration and landfilling. It is fed by the less frequent but much broader economic, social and demographic indicators of statistical offices.

According to the authors, without this separation decision support cannot work at the expected level, because the two task types require different model structures. The stakes of long-term decisions are well illustrated by our analysis of biodegradable waste landfilling and the extension of the EU ETS: shifts in the cost balance between landfilling and incineration shape treatment structures for decades.

Municipal solid waste prediction models: neural networks lead the field

The researchers screened publications from 2003 to 2023 indexed in Scopus up to 15 December 2023, following the PRISMA methodology. They only included studies that forecast waste quantity, describe the model structure and clearly present the data used. In the final set of 78 studies they identified 140 predictive applications, traceable to 36 distinct methods.

Machine learning methods account for about 73% of all applications. Within this group there are 39 supervised and 6 unsupervised learning tools, 46 deep learning applications and 11 ensemble methods. Neural networks and deep learning together represent 32.9% of applications, time-series forecasting 15.7% and system dynamics models 4.3%.

At study level, the ranking is as follows:

  • Artificial neural networks (ANN): 27 applications, 19.3% of studies – by far the most common method.
  • Support vector machines (SVM) and linear regression: 7.1% each.
  • ARIMA: 5.7%.
  • LSTM: 5.0%.
  • SARIMA: 4.3%.
  • Multivariate adaptive regression (MARS), random forest, decision tree, grey model and gradient boosting: 3.6% each.

Popularity, however, is not the same as suitability. The study summarises the strengths and weaknesses of the most common methods. Linear regression is simple and interpretable but sensitive to outliers. Decision trees are easy to read but prone to overfitting and ignore correlation between variables. Random forest reduces model variance, while SVM performs well on small samples but is sensitive to missing data and kernel selection.

Gradient boosting can rank feature importance but is difficult to train in parallel. Neural networks can describe any nonlinear relationship and are robust, but they need many parameters and are hard to interpret. ARIMA is simple and effective for stationary data but cannot capture nonlinear relationships. SARIMA handles seasonality at the cost of more parameters to estimate. LSTM is strong on sequence data but computationally demanding.

Data volume and number of variables: how the decision tree works

The most practical output of the study is a decision tree that leads to a recommended model along two questions. The first level is the number of variables: few (fewer than 3), moderate (3–7) or many (more than 7). The second level is the number of observations: sparse (fewer than 22), medium-term (22–162) or long-term (more than 162). The thresholds were not set arbitrarily: the authors trained a decision tree classifier on the model data of the reviewed studies and used the resulting cut-off points.

The proposed workflow is step by step. First, check whether the available number of observations reaches the minimum required by a given method. If it does, examine the set of available variables. If both conditions are met, build the model, and finally run cross-validation on a subset of the observations and check the results statistically.

According to the authors’ general guidance, when data is scarce, simple statistical tools – averaging, linear trends – analogy-based estimation drawing on regions with similar characteristics, or estimates based on local expert knowledge are the options. With a medium amount of data, time-series models such as ARIMA or regression models can be applied.

Large datasets open the way to multivariate regression, neural networks, decision trees and geostatistical methods that capture spatial and temporal patterns together. An important warning: on small datasets, models easily overfit, performing well on training data and poorly on new data. Support vector machines, random forest and gradient boosted decision trees are better choices for small samples.

The data series of the reviewed studies are revealing too. Most data comes from daily collection: the shortest daily series covers 90 days, the longest 28 years (Hong Kong). Weekly series span 4–12 years and monthly series 3–22 years. Among annual series the longest covers 49 years, and the number of observations is often increased by adding regions: the largest such dataset contains 15 years of data from 285 cities. Annual models used between 5 and 81 variables.

Economy, demography, weather: 455 variables in six groups

The 78 studies used a total of 455 explanatory (proxy) variables. After data cleaning, the researchers grouped them into six sets:

  • Economics – 140 variables (31%): national or regional GDP, tax revenue, exports of goods and services, number of companies.
  • Demography – 100 variables (22%): population, population density, age groups, number of households, family size.
  • Waste – 87 variables (19%): past waste quantities, waste per capita and fractions such as inert, glass, textile and plastic waste.
  • Social – 67 variables (15%): employment, unemployment, education by gender, urban poverty.
  • Geography – 41 variables (9%): annual mean temperature, rainy days, wind speed, solar radiation.
  • Welfare – 20 variables (4%): street cleaning area, number of hotel guests, urban residents’ spending on vegetables.

The most frequently used variables are past waste generation, population and waste fraction data, followed by GDP, population density, household income, education and the unemployment rate. The first three groups together make up 72% of the full variable set.

Practice confirms the weight of socio-demographic factors. In our article on Austrian waste management data for 2024, we showed that municipal waste per capita was 570 kg in Vienna and 951 kg in Upper Austria – a gap the report attributes to income, household size, commuting and tourism intensity. The EU-level picture is similar: according to Eurostat’s 2024 data, municipal waste per capita averaged 517 kg, with multiple-fold differences between member states. A well-chosen variable set is what allows a model to capture these differences.

Operations or strategy: two separate modelling worlds

The authors place the reviewed models in a “modelling problem space” defined by the number of observations and the number of variables, and two modelling worlds emerge.

Models built on many observations and few variables suit technological and operational tasks: route optimisation, scheduling, facility utilisation and capacity management. Their raw material is daily-resolution data, such as container fill-level measurements or collection vehicle logs. The same logic applies inside sorting plants, where systems built on real-time sensor data already optimise daily operations – as we described in our article on CP Group’s new AI-based waste sorting system.

Models built on many variables, by contrast, serve strategic planning: decarbonisation roadmaps, infrastructure investment and assessments of compliance with waste management directives. These typically rely on annual data and combine economic, demographic and geographic factors.

Between the two lies a “hybrid” region, where both the number of variables and the number of observations are large enough for the model to serve both purposes. Models with few observations and few variables, however, have limited predictive power: they are better suited to comparative assessment, scenario testing or early-stage estimates than to accurate forecasting.

Data quality is not only a modellers’ problem either. As we saw with the recently launched Packaging Data Hub, the gap between measured and reported values can distort policy decisions as well.

Accuracy on paper and in practice

Of the reviewed studies, 44 reported the coefficient of determination (R²), with values between 0.53 and 0.99996; 37 reported the root mean square error (RMSE) and 25 the mean absolute percentage error (MAPE). Only two studies reported accuracy.

The authors argue these metrics should be treated with caution. R² does not reveal systematic under- or overestimation and is dataset-dependent, so it is not always comparable across studies. RMSE expresses deviation in absolute units, which makes it hard to interpret without the scale of the target variable. MAPE is scale-independent, but several studies do not state clearly whether it is given as a ratio (for example 0.12) or as a percentage (12%) – which explains the reported range of 0.008 to 25.7, in reality not comparable.

For classification tasks, the authors recommend precision, recall and the F1-score alongside accuracy, and more generally k-fold cross-validation, which gives a more stable and less biased estimate of a model’s generalisation ability. One important limitation is stated by the authors themselves: the proposed framework does not evaluate predictive accuracy; it recommends which model structure fits a given analysis task and data situation.

The open question: forecasting composition

The study focuses on waste quantity, and the authors openly state that forecasting changes in composition remains an unsolved research task. Yet this is exactly what would allow necessary adjustments to treatment technologies to be planned in advance.

The main obstacle is data. National, annual datasets may be enough for a national waste management strategy, but technological development and implementation must account for local characteristics and seasonality – which requires monthly or higher-resolution data at the level of cities or collection districts. According to the authors, the most appropriate variable sets are likely to differ between countries and regions, so model identification and interpretation cannot be meaningfully carried out without domain expert input.

A further priority they identify is quantifying the system-level effects of a changing regulatory environment. This requires linking forecasting outputs with technological and infrastructure information and modelling individual waste fractions separately.

What waste management practitioners can take from this

The study comes from a Hungarian research group: it was produced within a project supported by Hungary’s National Research, Development and Innovation Office to establish a waste management competence centre at the University of Pannonia, and Viktor Sebestyén’s contribution was supported by the János Bolyai Research Scholarship of the Hungarian Academy of Sciences.

According to the paper, many of the relevant data sources are publicly available. Eurostat offers 4,243 data sources by statistical topic, including waste generation data, making it suitable for supervised learning tasks. The OECD provides 288 indicators in annual breakdown, the UN Sustainable Development Goals database contains 706 data series, and the World Bank offers useful proxy variables across 86 databases. Much of this can be queried dynamically via application programming interfaces (APIs).

The practical message for operators is twofold. For operational decisions – collection routes, sorting and treatment capacity – waste management companies’ own time series per fraction and per vehicle are the most valuable raw material, so collecting them systematically is a competitive advantage in itself. Strategic planning, in turn, needs a broad range of statistical and economic indicators paired with waste data.

The first step before choosing a model is therefore not picking the most fashionable algorithm, but an honest data inventory: how many observations and how many reliable variables are actually available. The University of Pannonia decision tree provides a starting point – and, as the authors recommend, the applicability of new models should be tested continuously, because machine learning tools are evolving at remarkable speed.


Municipal solid waste prediction – frequently asked questions

What is the purpose of municipal solid waste prediction?

The purpose of municipal solid waste prediction is to align the planning of collection, sorting and treatment with expected quantities. In the short term it supports optimising routes, scheduling and plant capacity; in the long term it informs infrastructure investment, decarbonisation roadmaps and compliance assessments. The University of Pannonia review treats these two task types as separate modelling worlds.

Which model is most often used to forecast waste quantities?

The most common method is the artificial neural network (ANN): 27 of the 140 model applications in the 78 reviewed studies were ANNs, appearing in 19.3% of studies. Support vector machines and linear regression follow at 7.1% each, then ARIMA at 5.7%, LSTM at 5.0% and SARIMA at 4.3%. Machine learning accounts for roughly 73% of all applications.

How much data does a reliable forecasting model need?

It depends on the chosen model. The researchers’ decision tree distinguishes sparse (fewer than 22), medium-term (22–162) and long-term (more than 162) observation series. With little data, averaging, linear trends, analogy or expert estimates are options; long time series also allow time-series and machine learning models. On small datasets overfitting is the main risk, so cross-validation is recommended.

Which variables can be used to forecast waste quantities?

The 78 studies used 455 variables in total, which the researchers grouped into six sets: economic (140), demographic (100), waste-related (87), social (67), geographic (41) and welfare (20) indicators. The most common are past waste generation, population, waste fractions, GDP, population density, household income, education and unemployment.

Can waste composition be forecast as well?

According to the authors, this remains an open research task. The proposed framework is designed for forecasting quantities. Forecasting changes in composition, which would help adjust treatment technologies in advance, would require at least monthly data at the level of cities or collection districts, and the appropriate variable set may also differ from country to country.


🔗 Official sources and references: 👉

Fejes R., Kurdi R., Sebestyén V.: Model structures and approaches for municipal solid waste prediction – A review. Resources, Conservation & Recycling Advances, 29 (2026), 200310. DOI: 10.1016/j.rcradv.2025.200310 – open access, CC BY 4.0 licence.

Consistent structure, quick answers: every article is built so that both readers and AI search tools can find the key points easily. Our content is produced with the help of AI tools and reviewed by our editor.

Ladányi Roland
Ladányi Rolandhttp://envilove.hu
Roland Ladányi is an environmental professional and waste management expert dedicated to promoting sustainability and the circular economy. As the founder and driving force behind the dontwasteit.hu platform, he provides up-to-date news, in-depth analysis, and practical solutions aimed at shaping an environmentally conscious mindset. His work focuses on waste reduction and efficient resource management, bridging the gap between technical expertise and clear, accessible public communication.
OLVASS TOVÁBB